AI + Pathway Intelligence
Current data. Connected evidence.
A clearer way forward.
See how AWS Lambda, vector search, Neo4j, TimesFM, LLMs and data science connect market evidence to your goals. We are also training an industry-specific model for education, skills and employment.
For your future, your team and the people you teach.
AI supports the decision. You remain in control.
“What makes sense for my next step?”
Your goal + connected evidence + decision rules
Architecture illustration · not a live market feed or prediction.
01How it connects
Pathway Intelligence connects what you want, what you can already do and what a realistic next step requires. The components do different jobs; no single tool is the whole answer.
A course may look relevant but repeat learning you already have. A role may use familiar tools but require work you have not demonstrated. A growing market may still offer few options within your available hours. Connecting these differences is the purpose of the system.
The architecture brings evidence to the decision, rather than asking a language model to invent the missing pieces. The sequence below explains the flow of information. It is not a questionnaire every reader has to complete, and not every question needs a forecast or graph search.
Capture
Market releases, job advertisements, course details and skills information arrive in different formats and at different times. Data ingestion brings authorised feeds, imports and permitted collection into a common process, retaining the source and the period each record describes.
Source records with collection dates, geographic scope and usage conditions.
You can ask what the information is based on, rather than rely only on an AI answer.
A recent retrieval does not make an old underlying statistic current. Coverage is source-dependent.
Prepare
Names, course identifiers, occupations, places and time periods need consistent meaning. Cleaning, duplicate detection, entity resolution and validation separate useful records from conflicts or incomplete data. The warehouse preserves the evidence; analytical features organise it for specific questions.
Consistent records, source history and quality flags that downstream analysis can inspect.
Two names for the same course should not become two options, and two different qualifications should not be silently merged.
An AI-proposed match between records remains a proposal until the relevant checks support it.
Connect
The graph layer, using Neo4j, represents connections between learning, tasks, skills, tools, roles and organisations. A useful path identifies what each relationship means and which evidence supports it. Search can then follow several connections instead of treating every record as an isolated result.
Traceable relationship paths, alongside their source, context and verification state.
You can explore how an activity relates to a work task, or how experience could connect to a different role.
A graph connection is not proof of personal suitability, formal credit or demonstrated competence.
Analyse
Data science examines comparable groups, recurring requirements and changes over time. Graph analysis can surface related capabilities; forecasting can estimate future values in suitable historical series. Each method answers a bounded question rather than replacing judgement about the whole decision.
Observed patterns, candidate relationships and separately labelled forecasts or estimates.
Your discussion can consider what is changing around a role, location or learning need.
Association is not causation. A national trend cannot establish an individual opportunity.
Compare
Relevant options need to be checked against your goal, existing evidence, required entry conditions and practical constraints. The decision layer can distinguish an option to explore from a requirement to verify, and a learning need from a need for experience or better evidence.
A bounded comparison with reasons, trade-offs and important unknowns.
The next step may be learning, recognition, work, support or further exploration, not automatically a course.
Search similarity, market demand and forecast confidence are not your probability of being hired.
Explain
Language models help interpret questions and explain the relevant evidence in ordinary language. Oala carries the conversation; the workspace can show a comparison, relationship, project or pathway. Both need the same current understanding of your goal and corrections.
An explanation you can question, with an interactive view when it adds value.
You can see why an option matters and change an assumption without starting the discussion again.
Retrieved context can improve grounding, but it does not make generated text infallible.
Progress
Understanding and action are different steps. A request, application, provider contact or shared record needs a clear destination and your approval. Where outcomes can be collected appropriately, they can inform evaluation and a later review of the pathway.
An authorised next action, or a clear question to resolve before acting.
The system is intended to help with the next step after the recommendation, not leave you with a list.
Selecting an option is not consent. An application, interview and sustained employment are different outcomes.
A useful answer may be a comparison, a skills question, a learning plan or a next action to consider. It may also be: this requirement needs checking before a responsible recommendation can be made.
Architecture / four connected views
Explore the data plane, retrieval path and training lifecycle. Each view names the technical responsibility, what moves between components and where control stays outside a model.
Logical architecture, not a live deployment map. Lambda and vector search are in use; industry-model training is underway. The diagrams make the intended connections explicit without asserting a particular queue, training stack, cloud topology or measured performance.
Data collection, evidence stores, analysis and request execution are distinct layers. The named technologies work together; no model is the entire product.
On a small screen, scroll the diagram horizontally or read its text explanation below.
The pipeline retains original evidence, reconciles identity, versions changes and updates derived indexes. Failed records do not silently become trusted data.
On a small screen, scroll the diagram horizontally or read its text explanation below.
Semantic relevance is one input, not an eligibility decision. Exact lookups, graph context, policy checks and source evidence remain visible in the response path.
On a small screen, scroll the diagram horizontally or read its text explanation below.
Our industry-model training work is separate from live retrieval. A model release needs versioned data, evaluation, approval and a rollback path.
On a small screen, scroll the diagram horizontally or read its text explanation below.
Serverless execution / AWS Lambda
Yuzee uses Lambda for event-driven processing and matching work. It is the execution layer for bounded tasks, not the database, the knowledge graph or the model itself.
AWS Lambda runs code in response to events or API requests while AWS manages the underlying compute. In Yuzee, that fits collection, validation, enrichment and retrieval orchestration. The result can be a validated record, a candidate set or a job handed to another service. [13]
Why it matters to you: incoming information and interactive requests can be processed as separate, bounded jobs rather than tying every task to one large model call.
Long-running collection, large graph computations and model training belong on compute selected for their duration, memory and accelerator needs. A Lambda function can initiate or check those jobs without running the entire training process inside an interactive request.
For standard Lambda function invocations, the documented maximum duration is 15 minutes. This is a platform limit, not a Yuzee response-time promise. [13]
A repeated event must not create a second record, application or downstream action. Use a stable event identity and explicit write semantics.
Limit concurrent work to what databases, indexes and model endpoints can sustain. A queue buffers bursts; retries have limits and failed work has a review path.
Carry only the request's permitted context. Do not use a warm function's memory as a durable user profile or a cross-user conversation store.
Associate the request with processing, retrieval and model versions. Trace failures without copying confidential content into unrestricted logs.
Reference reliability controls, informed by AWS guidance on idempotent processing, restricted permissions, concurrency and queue retries. These are not claims that every control has been independently audited in Yuzee. [14] [15]
02The information behind the answer
Yuzee captures and organises market, job, skills and education information. Its value is not a promise to know everything; it is the ability to connect relevant records and retain where they came from.
Advertised tasks, experience, tools, working conditions and locations, alongside official labour-market series and occupation information. [1]
Capability definitions, occupational tasks, software and equipment references, and the evidence linking them. Tool familiarity stays separate from the ability to perform a task. [2]
Courses, units, learning outcomes, delivery, entry requirements, projects and assessments where these are published or supplied. Missing detail remains visible. [3]
Employer skills briefs, sector context and business information where available and permitted. A particular employer's configuration is not treated as an industry-wide rule.
Location, regional context, affordability, delivery arrangements and support or eligibility information. A pathway must be practical for the person, not merely possible on paper. [4]
User corrections and, where collected with appropriate permissions, information about learning, requests and later outcomes. Private records need their own access controls.
Freshness with context
A webpage collected today can describe last year's outcomes. An official monthly release is not a live vacancy list. A provider may update an intake without updating every page that mentions it.
Our data standard is to retain source, time and coverage information, flag uncertainty and recheck time-sensitive conditions before consequential action. Refresh frequency depends on the source, permission and intended use.
For example, JSA's Internet Vacancy Index measures online advertisements within a defined source set. It does not represent every vacancy or every way employers recruit. A removed advertisement also does not, by itself, establish that someone was hired. [1]
03Neo4j & connected evidence
A graph makes the connections between records explicit. A useful connection has a meaning, context and supporting evidence, not just two similar words.
Solid: example structural links. Dashed: suggested mappings. All are illustrative, not verified records or a live Neo4j query.
Neo4j is the relationship layer, not the source of truth by itself. Its property-graph model stores entities, named connections and their attributes. Yuzee's architecture uses that structure to follow evidence across learning and work while preserving the link back to source records. [5]
Connecting a course to a task does not prove that everyone completing the course can perform that task. A suggested skill-to-role link needs review. Private user evidence remains permissioned. The graph should make the explanation easier to inspect, not make uncertainty disappear.
Vector search / hybrid retrieval / GraphRAG
Vector search helps retrieve descriptions that express related ideas even when their wording differs. It supplies candidates for a decision; it is not a verdict about a person.
An embedding turns a description into a numerical vector. Approximate nearest-neighbour (ANN) search retrieves nearby vectors efficiently; exact nearest-neighbour search evaluates the indexed vectors exhaustively. MongoDB Atlas Vector Search appears in Yuzee's documented course-matching pipeline. [16]
The query and indexed records must use compatible representations: the same embedding model and expected dimensions, with a versioned preprocessing contract. Equal vector length alone is not enough to make two models compatible. [17]
A request about “handling customer complaints” may be semantically related to “service recovery”. That can surface a useful course, project or task to inspect. It does not establish proficiency, equivalent qualifications or employer acceptance.
Exact identifiers and verified filters still matter. An occupation, credential code, region or entry requirement must not disappear because another result is closer in embedding space.
Keyword scoring and vector retrieval can be combined into one ranked evidence set. Reciprocal rank fusion (RRF) is one method: it combines ranks without pretending keyword scores and vector distances share the same scale. This is a reference design choice, not a published Yuzee weighting formula. [7]
A cross-encoder can then evaluate a query and candidate together for relevance. It remains a re-ranker, not an assessor of the user's competence. [20]
Graph-augmented retrieval can connect a retrieved course to its units, tasks, requirements or tools before selected evidence reaches the LLM. Neo4j documents vector, hybrid and Cypher-augmented retrieval patterns. [8]
For this reference architecture, graph traversal is bounded and queries are allow-listed or validated. A model-proposed query is not unrestricted database authority; access checks happen before content reaches the model.
Representation: embedding-model ID, revision, dimensions, normalisation, chunking and query/document prefixes. Changing a model requires a controlled index migration or re-embedding plan.
Record identity: canonical entity ID, source revision, record status and access scope. Merging duplicate search results must not merge different qualifications or different people.
Evidence: source passage, observed date, validity scope and relationship type. A source refresh can invalidate a cached comparison or graph link.
Output: candidate IDs with retrieval scores and provenance. Suitability reasons, missing information and required verification belong to a later decision step.
04The intelligence layer
Data engineering, graph analysis, search, forecasting, statistics and language models complement each other. The right method depends on the question, the available evidence and the cost of being wrong.
Collection, cleaning, versioning and quality checks give every later component a reliable starting point. Separating original records from derived analysis makes a correction easier to trace and apply.
Ingestion workers, canonical identifiers, duplicate checks, source lineage and versioned analytical features. Refresh and expiry policies are set by source and use case.
Neo4j stores entities and named relationships. Its Graph Data Science tools include similarity, communities and path-finding. In this architecture, these methods can help investigate related capabilities and possible routes. [5][6]
Typed, directional relationships; evidence attached to the connection; bounded traversal; separate access controls for personal data. Graph analysis proposes patterns that still need domain checks.
Exact identifiers and keyword search preserve specific meaning. Embeddings help find related descriptions. Graph retrieval adds connected context. Combining these approaches can produce a candidate evidence set for further checking. [7][8]
Full-text and vector retrieval, controlled graph queries, fusion and re-ranking. Retrieval-augmented generation, or RAG, gives a language model selected evidence before it explains an answer.
TimesFM is Google Research's pretrained time-series forecasting model. Its role here is to estimate future values for suitable historical series, such as a defined job-advertisement index, with uncertainty considered alongside the estimate. [9][10]
Time-indexed observations, a named target and horizon, stored model version, point and quantile forecasts, and evaluation against simple baselines. Use only series that pass relevant quality checks.
Cohort comparisons, change detection, clustering and carefully evaluated models can help distinguish recurring signals from noise. Constrained comparisons can investigate feasible learning and work routes without hiding the trade-offs.
Retain definitions and denominators; inspect missing data, subgroup performance and drift; avoid future-data leakage; use experiments or stronger study designs for causal claims. These are analytical methods, not a claim that every model is deployed.
Large language models help interpret a question, extract candidate concepts, explain selected evidence and support guided learning. Orchestration assigns those tasks to appropriate components instead of asking one model to invent records and decide everything. [8][11]
Shared, scoped context; structured outputs; authorised tools; evidence checks; fallbacks; and explicit permission before external actions. The database and access layer remain separate from generated language.
Serverless functions run collection, processing and matching tasks. They coordinate with databases, indexes and model services rather than replacing them. [13]
Event contracts, scoped permissions, safe retries and separation of online requests from offline compute.
What this does not establish
A serverless component does not imply unlimited capacity or an end-to-end latency guarantee.
Our model-training work targets education, skills and employment. The purpose is to improve domain interpretation and evidence-aware assistance alongside retrieval. [18]
Task-specific training, reviewed data, held-out evaluation and versioned release. The exact model recipe is not asserted on this page.
What this does not establish
Training in progress is not a released benchmark result. Fresh facts still need retrieval.
Scheduled and event-driven work can update records, build search indexes, refresh graph views and evaluate forecast series. Heavy data processing should not have to happen afresh for every conversation. The frequency and coverage of those jobs remain specific to each integration.
At question time, authorised services select relevant evidence and permitted actions. The warehouse, search index and graph view are different components with different jobs. A language model should not receive unrestricted database access or an entire private history merely to explain one comparison.
It brings requirements, evidence and constraints together. A high search score cannot waive an entry condition. A rising forecast cannot override your availability. A predicted graph link cannot certify a skill. Exact ordering and commercial-treatment rules belong in the Yuzee Methodology .
Our industry-specific model / training in progress
We are training our own model for education, skills and employment. The objective is domain-specific understanding: how goals, capabilities, requirements, learning and work fit together.
A specialised model does not replace current evidence. It should help interpret the domain and support better decisions, while live fees, entry conditions, vacancies and source changes still come through retrieval.
A useful industry model should distinguish a skill from a tool, a preference from a requirement, a course outcome from personal evidence, and an unknown capability from a confirmed gap. It should understand a change in goal without inventing additional experience.
Relevant tasks include classifying the user's current objective, mapping terminology, identifying candidate relationships, evaluating relevance and generating evidence-bounded explanations. These are development objectives to evaluate separately, not a claim that one released model already performs them all.
It does not automatically mean a new foundation model trained from scratch. A domain model may build on existing model architectures and authorised data. The base model, training method, corpus size and validated benchmark results need a version-specific model card before stronger claims are made.
Prompt engineering, updating a graph and calibrating retrieval thresholds are useful engineering work, but are not the same as changing model weights. We keep these layers separate.
Train retrieval representations on reviewed pairs or examples relevant to the domain. Evaluate whether suitable evidence is retrieved despite different terminology. [19]
Candidate method; not a declared training recipe.Learn which candidate best addresses the request using judged query-result examples, including plausible but incorrect matches. Keep personal suitability separate. [20]
Candidate method; requires representative labels.Train on reviewed examples of the intended response or task. Parameter-efficient methods such as LoRA adapt selected parameters instead of updating every weight. [18] [21]
Candidate method; no base model or adapter is asserted here.Separate the data. Group related records and repeated examples before splitting; keep held-out tasks and later periods out of tuning.
Measure the right failure. A lower training loss does not establish retrieval quality, sound advice, safe actions or useful learning.
Preserve lineage. Record data rights, labels, model revision, training configuration, evaluations and approved release scope.
Control change. Promote a version only after review. Monitor it and preserve rollback; user feedback is not automatic permission to retrain.
This page describes the training work and reference lifecycle. It does not publish a parameter count, claim a proprietary foundation model, or report unverified improvements. Specialist training remains subject to evidence, privacy and release controls.
Explore the training architecture →05TimesFM & time-based signals
Forecasting can inform planning when the historical series is suitable. The output must remain separate from observed data and from the recommendation for a particular person or organisation.
Synthetic illustration · not market data, a TimesFM run or a calibrated uncertainty interval.
For a clearly defined, sufficiently supported series, compare forecasts with simple baselines and held-out observations. Show the horizon, source period, model version and evaluated uncertainty.
The band is an illustration of uncertainty, not a calibrated interval or a current Yuzee prediction.
A forecast might concern a defined regional advertisement index or a consistent series of course enquiries. It needs a time interval, a reference period and enough usable history. It cannot turn a broad question such as “Will I succeed in this career?” into a reliable personal prediction.
TimesFM is the forecasting component, not the recommendation engine. Its point and quantile outputs need evaluation on the series and horizon in question; a pretrained model does not remove that responsibility. [9][10]
What a defined source recorded in a specified period. It may be revised later.
A model-based estimate of later values, conditional on the available history and assumptions.
An interpretation of relevant evidence alongside goals, requirements and practical choices. The forecast is only one input.
06The system in your situation
Select a perspective, then change one detail. See why a useful next step should change when the relevant evidence or circumstances change.
Illustrative walkthroughs. These controls do not call live AI, search candidates, deliver training or query the Yuzee warehouse.
Let's compare the work you have actually done with the tasks in an operations role. Some experience may be relevant; other requirements may need evidence, practice or verification before choosing a course.
A new fact should update both the explanation and the comparison, without being treated as proof of additional skills.
The graph supplies possible connections. The person's evidence and constraints determine what is worth investigating. The language model explains the difference.
Start with your service rules, tools and the level of independence expected. Then compare existing evidence with the tasks. The response may be recruitment, new-starter preparation, staff development or a combination.
A new fact should update both the explanation and the comparison, without being treated as proof of additional skills.
Employer demand becomes a defined task and evidence requirement, not just another keyword used to filter people.
First check whether this is a recurring task need or a small number of product mentions. Compare it with your existing foundations, practice and assessment before deciding what to update.
A new fact should update both the explanation and the comparison, without being treated as proof of additional skills.
Data science identifies a question to investigate. Graph relationships connect it to course evidence. Educators retain curriculum and assessment responsibility.
Which reporting tasks have you performed, and what does your intended work require? The useful next step depends on the workflow, the underlying knowledge and what you can already demonstrate, not just the popularity of a product.
A new fact should update both the explanation and the comparison, without being treated as proof of additional skills.
A tool can improve an output without proving the person has acquired all the underlying skills.
The conversation should explain the immediate question, use relevant context and ask only what materially improves the next decision.
Compare options, follow relationships and inspect evidence. It complements the conversation; it is not a second independent counsellor.
Exploring, comparing and preparing are not permission to apply, enrol or share a record. The next action needs a clear service and approval.
07From information to capability
The learning direction is not simply to generate more content. It is to connect a defined need with the knowledge, practice, feedback and evidence that could support the next step.
For a learner, a project should have a reason: a task to understand, a problem to solve and a way to review the work. AI can support explanations, safe practice scenarios and reflection around appropriate materials. It should not quietly do all the thinking and then treat the resulting output as the learner's independent ability.
For an employer or educator, the important questions include who sets the task, who reviews it and what the evidence actually supports. Foundational learning, project work and workplace experience can complement one another without being treated as interchangeable. Human-centred design is also central to UNESCO's guidance on educational AI. [12]
Specify the capability, context and expected standard. Check what the person already knows.
Use appropriate material to explain concepts, methods, risks and why the work is done this way.
Practise with a meaningful brief and safe example data. Make the learner's contribution visible.
Examine decisions, accuracy and changes to the work. Identify who provides feedback and assessment.
Revisit what needs practice, supervision or formal verification before increasing responsibility.
Define the tasks your team needs to perform, what matters on entry and what can be developed with support. That gives a clearer basis for recruiting, preparing new starters, developing existing staff or combining these routes.
The connected-learning direction is to relate that brief to existing evidence, learning resources, projects and provider responses. Faster hiring and stronger performance are aims to evaluate, not guarantees attached to a match.
Compare recurring task needs with your existing curriculum, projects and assessments. A product change may need updated practice; a deeper change may need new knowledge. Missing public information should not be misread as missing teaching.
The aim is to help relevant learners and employers recognise what you offer, and identify useful extensions where needed. Educators retain responsibility for curriculum, teaching and assessment; demand signals inform rather than dictate the response.
Learning delivery, approved materials, provider capacity, assessment arrangements and access depend on the programme and available services. A simulated task is not the same as independent workplace performance. The complete learning principles sit in Our methodology .
08Data science & evaluation
A working data pipeline is not the same as a correct recommendation. A convincing explanation is not the same as useful learning. The evaluation needs to match the claim.
Do records retain scope, dates and rights? Are duplicates, changes and missing fields handled correctly?
A large warehouse is not evidence of complete or representative coverage.
Do sampled links mean what they claim, and does the cited material support each connection?
A plausible graph path or predicted edge is not a confirmed pathway.
Does the system retrieve the right evidence, preserve corrections and keep generated claims within the evidence?
A valid response format does not prove the advice is correct.
Does the model perform acceptably for the target series and horizon, compared with baselines and later observations?
One benchmark or successful run does not validate every forecast.
Can the person explain their choice, demonstrate a task and identify what still requires practice or review?
AI-assisted output alone does not prove independent understanding.
What happened after the next step, for whom and under which conditions? How are errors, selection effects and unequal access examined?
Clicks, completions and hires are different measures. Association does not establish that Yuzee caused an outcome.
These are evaluation responsibilities, not claims that every test has been completed. New data, reviewed corrections and suitable outcome evidence should trigger reassessment. Model or policy changes need versioning and checks before release, not automatic self-improvement from every click. [11]
MLOps / reliability / security
The reference operating model treats data, indexes, prompts, policies and models as versioned dependencies. A change in one layer can change the answer, even when the interface looks the same.
Monitor source coverage, failed imports, duplicate rates and index lag. Carry observed time and validity scope through retrieval. Do not call a newly fetched old statistic a live market signal.
Record model, dataset, feature and configuration versions. Compare a candidate release with the current baseline; use controlled rollout and rollback rather than silently replacing behaviour.
Keep a request ID, scoped context revision, retrieved source IDs, tool permissions and validation results. Inspect grounding and correction handling, not only whether the JSON was valid.
Apply user and organisation permissions before retrieval and again before an action. Validate tool names and arguments. Treat retrieved documents as data, never as higher-priority instructions.
Retrieval: Recall@k measures how much of the relevant judged evidence appears in the first k results; nDCG@k measures ranking quality against graded relevance. State the dataset and judge agreement.
Forecasting: use chronological backtests, MAE or suitable scale-aware errors, and pinball loss or interval coverage for quantile forecasts. Report the target, horizon and baseline; sparse or inconsistent series can warrant abstention.
Runtime: separate cold and warm latency, p50 and p95, error rates, queue delay and cost per completed task. A fast component is not proof of a fast end-to-end journey.
Safety and usefulness: inspect invented claims, unsupported certainty, unauthorised actions, subgroup errors and whether people understand the next step. Publish measured results with scope; no performance figures are claimed here. [11]
09Evidence & your control
Connections and models can help organise a difficult decision. They should not hide the evidence, replace your choice or make a weak claim look certain.
The ambition is to expand access to useful guidance, relevant learning and opportunity. Better economic participation is a contribution to work towards and measure, not a person's worth or an outcome this page claims to have proven.
A provider record, a person's account, an AI-extracted relationship and a forecast need different labels. Each should be shown with enough context to understand what it supports. “Unknown” must not silently become “missing”, and confidence in retrieval must not become confidence in someone's ability.
Relevant user information and corrections can support a decision without being exposed to every model, employer or institution. Collection, retention, sharing and reuse require appropriate permissions and stated purposes. Research or model-improvement use needs its own governance; using a service is not blanket consent.
A selected course, comparison or prepared plan does not authorise contact, an application or enrolment. Before acting, the destination, information being shared, service and consequences should be clear. Formal assessment, admission, hiring and regulatory decisions stay with the appropriate people and organisations.
Retrieved pages and job advertisements are data, not instructions that can override the system's rules. Approved queries and tool permissions limit what an AI component can access or change. When retrieval fails or sources disagree, the system should preserve the gap and offer a useful next check rather than fabricate a result.
The guiding standard is that a commercial relationship should not silently determine an organic recommendation. Sponsorship, restricted catalogue coverage and available service partners need clear treatment. A no-new-course option must remain possible when that is the appropriate next step. See Methodology for the applicable principles and disclosures.
10The development direction
This page explains the architecture and why its components belong together. It is not a real-time service inventory, a deployment audit or a claim that every example is available today.
Yuzee uses Lambda and vector search alongside data capture, warehouse work and its graph and model layers. Neo4j, TimesFM, language models and data science have distinct roles in the architecture.
These are named technology choices, not a claim that all diagrams describe a fully released deployment. Current feature scope, source coverage and measured performance need release-specific evidence.
Industry-specific model training is in progress. In parallel, the work is to connect retrieval, graph context, decision rules, corrections and learning experiences into a dependable whole. A training run, an index update and a product release are separate events.
The priority is a dependable end-to-end decision, not the largest number of models, tools or records.
Deeper employer skills requests, AI-supported practice, provider responses and project evidence could extend the pathway after the first recommendation. Permission-based partner integrations, analytical APIs and localised support could make the same approach useful in more settings.
Expansion depends on local evidence, access, teaching and assessment arrangements, partner agreements and outcome validation. No delivery date or universal coverage is promised here.
Work on something that matters
Reliable data engineering, thoughtful AI, learning design and careful evaluation belong in the same conversation. Bring the ability to make evidence useful, challenge assumptions and test whether a real person can take a better next step.
Explore the principles and ways to contributeHow do we make a useful relationship traceable?
When does practice become evidence someone can rely on?
How do we make uncertainty understandable?
How do we learn whether the next step actually helped?
An invitation to contribute, not a list of confirmed vacancies or a request for unpaid production work. Roles, scope, terms and available opportunities require a separate conversation.
Technical glossary
A reference for readers who want to inspect the architecture without guessing what each acronym means. These definitions describe concepts; they do not imply every technique is deployed.
30 terms
Event-driven compute that runs bounded code when an event or request arrives. It executes work; it is not a database or a language model.
A numerical representation of content used for similarity search. It captures aspects of meaning, not proof that a statement is true.
Approximate nearest-neighbour search finds nearby vectors without exhaustively scoring every vector. HNSW is one graph-based index technique; its links are not the domain knowledge graph.
A keyword relevance scoring method based on term matching and document statistics. It complements meaning-based retrieval rather than measuring personal suitability.
Combining different retrieval methods, such as keyword and vector search, so exact terminology and semantic relationships can both contribute candidates.
Reciprocal rank fusion combines ranked lists. It is a possible fusion method, not an undisclosed Yuzee suitability score.
A model that considers a query and a candidate together to score relevance. Usually applied to a smaller candidate set after retrieval.
A database model built around entities, named relationships and their properties. In this design, a relationship also needs evidence and scope.
Neo4j query language for finding and traversing graph patterns. Production access should be authorised and bounded; generated queries are not automatically trusted.
A taxonomy organises terms; an ontology also defines entity types and permissible relationships. Neither certifies the capability of an individual.
Determining whether records refer to the same real-world entity. Preserve ambiguous matches instead of merging records solely because names look similar.
The history connecting a result to its sources and transformations. It helps show which outputs may need revision after a source changes.
Processing triggered by new or changed data. A scheduled import is another trigger; neither guarantees that the underlying source is real-time.
Processing the same event again without creating an unintended duplicate effect. Important when delivery or retries repeat work.
Backpressure limits incoming work to sustainable capacity. A dead-letter path holds repeatedly failed work for diagnosis instead of endless retries.
Retrieval-augmented generation gives a language model selected external evidence. GraphRAG adds relationship context. Neither changes model weights by itself.
A large language model generates or interprets language. Its fluency is not a guarantee of factual accuracy, domain competence or permission to act.
A Google Research time-series foundation model. It forecasts numerical sequences; it does not decide which career is right for a person.
Supervised fine-tuning adapts model weights using reviewed training examples. A lower training loss does not by itself establish real-world quality.
Parameter-efficient fine-tuning changes a smaller set of parameters. LoRA uses low-rank updates. These are training options, not confirmed Yuzee implementation choices.
An example that looks relevant but is wrong for the task. Such examples can teach a model to distinguish a tool name from the capability or requirement being asked about.
A held-out test is reserved for evaluation. Leakage occurs when test information or future data influences training, making results look better than they generalise.
A registry tracks versioned model artefacts. A model card describes intended uses, training scope, evaluation and limitations so a release can be examined.
A change in inputs, relationships or performance after deployment. Source coverage, vocabulary or employer requirements can shift without a model-code change.
A forecast evaluation that repeatedly trains or conditions on earlier observations and tests later ones. It preserves the order in which information became available.
A forecast of a point in a predictive distribution, rather than a guaranteed boundary. Empirical coverage needs testing on the intended series.
Checking whether a score or uncertainty statement corresponds to observed outcomes. Tuning a routing threshold is not the same as validating a hiring probability.
The release, evaluation and monitoring practices around machine-learning and language-model systems. Tools support these practices; a label is not evidence they are effective.
Give a component only the access needed for the request and organisation. A shared index must not expose another user or employer's private records.
A controlled rollout exposes a new version to a limited scope. Rollback restores a known version if checks fail; it is not a substitute for pre-release evaluation.
No matching term. Try a shorter word or clear your search.
11Sources & scope
The references below establish technology capabilities, data definitions and relevant governance or learning principles. They do not endorse Yuzee or verify its implementation. Yuzee-specific connections, scenarios and evaluation criteria on this page describe the architecture and intended approach.
Reviewed 21 September 2026. Tool documentation changes; model, dataset and product versions should be recorded in the service that uses them.
What it supports: Defines the monthly online-advertisement series, occupational and geographic coding, coverage and conceptual limitations.
What it does not establish: Advertisements are not all vacancies; this does not establish Yuzee's coverage or an individual's prospects.
Read the original source (opens a new tab)What it supports: Describes the purpose of a shared, evidence-based language for skills across education and employment.
What it does not establish: A framework does not certify personal capability or prove that a particular Yuzee mapping is correct.
Read the original source (opens a new tab)What it supports: Describes the scope of nationally recognised VET activity in the collection.
What it does not establish: It is not a live inventory of every course, all non-accredited learning or individual learner proficiency.
Read the original source (opens a new tab)What it supports: Official guidance on accessing ABS data and metadata programmatically.
What it does not establish: API access is not proof that every dataset is integrated into Yuzee or refreshed continuously.
Read the original source (opens a new tab)What it supports: Explains nodes, relationships and properties, including directed, typed connections.
What it does not establish: Product documentation describes the technology, not the quality or deployment of Yuzee's graph.
Read the original source (opens a new tab)What it supports: Describes graph algorithms and machine-learning pipelines. Features have different maturity levels.
What it does not establish: Available algorithms are not evidence that Yuzee uses all of them or that a predicted relationship is valid.
Read the original source (opens a new tab)What it supports: Explains combining full-text and vector search into a ranked result set.
What it does not establish: Retrieval relevance does not establish correctness, eligibility or personal suitability.
Read the original source (opens a new tab)What it supports: Documents graph, vector and hybrid retrieval patterns that supply context to a language model.
What it does not establish: A retrieval pattern is not a guarantee against hallucination or an endorsement of this page.
Read the original source (opens a new tab)What it supports: Describes the TimesFM forecasting model family and version-dependent capabilities. Record the checkpoint and package used for each run rather than relying on an unpinned "latest" label.
What it does not establish: The package version and model version are different. Upstream capabilities do not establish accuracy on Yuzee data.
Read the original source (opens a new tab)What it supports: Explains the original forecasting-model approach and research context.
What it does not establish: Research benchmarks are not evidence of accuracy for a particular occupation, location, horizon or service.
Read the original source (opens a new tab)What it supports: Provides a framework for addressing AI risks across design, development, use and evaluation.
What it does not establish: Referencing the framework does not claim certification, compliance or independent validation of Yuzee.
Read the original source (opens a new tab)What it supports: Human-centred guidance on using generative AI in education and research, published in 2023.
What it does not establish: Guidance does not establish the effectiveness of a specific AI tutor, project or Yuzee learning intervention.
Read the original source (opens a new tab)What it supports: Event-driven function execution and the distinction between bounded functions and other compute patterns.
What it does not establish: A platform capability is not a Yuzee performance benchmark or deployment map.
Read the original source ↗ (opens a new tab)What it supports: Idempotency, restrictive permissions, concurrency controls and operational telemetry.
What it does not establish: Recommended controls are not evidence that Yuzee has completed a security or reliability audit.
Read the original source ↗ (opens a new tab)What it supports: Queue-triggered processing, repeated delivery and partial-batch failure handling.
What it does not establish: SQS is an illustrative queue integration here, not a declared replacement for existing Yuzee messaging.
Read the original source ↗ (opens a new tab)What it supports: Embeddings, semantic retrieval, ANN/HNSW, exact search and metadata filters.
What it does not establish: A vendor feature or similarity score does not establish candidate competence or Yuzee deployment coverage.
Read the original source ↗ (opens a new tab)What it supports: The same embedding model and index dimensions are required for a compatible query representation.
What it does not establish: No specific embedding version, dimension count or query API is prescribed for Yuzee production by this page.
Read the original source ↗ (opens a new tab)What it supports: Training language models on supervised examples and supported dataset forms.
What it does not establish: An available trainer is not proof of the base model, recipe or results of Yuzee's industry-model work.
Read the original source ↗ (opens a new tab)What it supports: Task-specific embedding-model training and evaluation concepts.
What it does not establish: An embedding-training option is not a confirmed current Yuzee training method.
Read the original source ↗ (opens a new tab)What it supports: Cross-encoder re-ranking and query-passage relevance training.
What it does not establish: Relevance ranking is not validated personal suitability or a hiring probability.
Read the original source ↗ (opens a new tab)What it supports: Low-rank adaptation as a parameter-efficient training approach.
What it does not establish: LoRA is a candidate technique explained here; its use by Yuzee is not asserted.
Read the original source ↗ (opens a new tab)About the examples: the graph, chart and conversations are explanatory illustrations, not live records, model outputs or observed Yuzee outcomes. This HTML makes no background network requests and collects no personal information. Links you choose to open lead to other pages or services. Public data coverage and operational availability must be checked separately.
AI supports the work. People give it purpose.
Connect what is known, make the unknown visible and keep the choice with the person it affects.
This will only take a moment.
