Architecture

1. 1De-identification pass over incoming notes
2. 2Structure-aware chunking by encounter section
3. 3Hybrid retrieval over the patient's own record only
4. 4Grounded drafting with per-claim citation
5. 5Clinician review queue with accept / amend / reject
6. 6Immutable audit log of every generated draft

Requirements that bind

• No claim without a citation to a source note
• Refuse and escalate when retrieval is thin, never interpolate
• PHI never leaves the tenant boundary
• Full audit trail of prompt, retrieval set and output

Stack

Ships with

Dataset

1,200 synthetic encounter notes across 9 specialties (no real PHI)

Deployment

Azure Container Apps, private networking, managed identity

Part of

Architecture

1. 1Feature retrieval from the transaction store
2. 2Agent plans an investigation over graph, device and merchant tools
3. 3Evidence table assembled with provenance per row
4. 4Narrative drafted against a fixed analyst template
5. 5Interrupt for analyst approval before case creation

Requirements that bind

• Every assertion traceable to a query and its result
• Deterministic replay of any completed investigation
• Hard cap on tool calls and spend per case
• Model output never writes to the case system unapproved

Stack

Ships with

Dataset

180k synthetic transactions with labelled fraud rings

Deployment

Kubernetes with KEDA scaling on case-queue depth

Part of

Architecture

1. 1Attribute extraction from the natural-language query
2. 2Hard filters applied before vector search, not after
3. 3Hybrid retrieval over product text and specifications
4. 4Cross-encoder rerank against the original intent
5. 5Streamed results with an explanation of the ranking

Requirements that bind

• Price, stock and location filters are absolute, never approximate
• P95 latency under 800ms on a 2M-SKU catalogue
• Zero hallucinated SKUs — every result exists in the index

Stack

Ships with

Dataset

2M synthetic SKUs with attributes, stock and regional pricing

Deployment

Docker Compose reference stack, Azure-ready

Part of

Architecture

1. 1Multi-format intake: PDF, photo, email, call transcript
2. 2Damage classification with confidence thresholds
3. 3Clause retrieval scoped to the claimant's own policy
4. 4Coverage reasoning with clause-level citation
5. 5Interrupt for adjuster approval above a value threshold
6. 6Checkpointed state so a claim can resume days later

Requirements that bind

• Coverage decisions cite the exact clause and version
• Anything above the threshold requires a human approval
• A killed process resumes with no duplicated side effects

Stack

Ships with

Dataset

6,500 synthetic FNOL packets and 40 policy wordings

Deployment

Azure Container Apps with Postgres checkpointing

Part of

Architecture

1. 1Redaction of name, gender, age and location before scoring
2. 2Rubric-driven extraction with evidence spans
3. 3Per-criterion scoring with quoted justification
4. 4Bias evaluation suite run on every prompt change
5. 5Recruiter review with one-click override and reason

Requirements that bind

• Redaction verified by test, not asserted by comment
• No score without a quoted evidence span
• Bias eval must pass in CI before a prompt ships
• Every automated decision is overridable and logged

Stack

Ships with

Dataset

4,000 synthetic applications with balanced attribute distributions

Deployment

Docker, with a GitHub Actions eval gate

Part of

Architecture

1. 1Streaming anomaly detection over sensor telemetry
2. 2Retrieval across manuals, work orders and prior incidents
3. 3Agent correlates signal, history and procedure
4. 4Ranked checklist with expected time and parts
5. 5Feedback loop from the completed work order

Requirements that bind

• Every recommendation cites a manual section or prior incident
• Runs on-premises — no telemetry leaves the plant network
• Degrades to rules-only when the model is unreachable

Stack

Ships with

Dataset

18 months of synthetic telemetry from 40 machines, plus 900 work orders

Deployment

On-prem Kubernetes, air-gapped model endpoint

Part of

Architecture

1. 1Clause segmentation with heading and numbering awareness
2. 2Clause-type classification against a fixed taxonomy
3. 3Nearest-standard retrieval per clause
4. 4Deviation diff with risk rating and rationale
5. 5Reviewer workspace with accept / negotiate / escalate

Requirements that bind

• Deviations shown as a literal diff, never paraphrased
• Unclassifiable clauses are flagged, not silently dropped
• Document never leaves the tenant's storage account

Stack

Ships with

Dataset

800 synthetic commercial contracts and 60 standard positions

Deployment

Azure Container Apps with private blob storage

Part of

Architecture

1. 1Misconception taxonomy per topic
2. 2Diagnostic questioning before any explanation
3. 3Explanation generated against the diagnosed gap
4. 4Spaced-repetition scheduling of weak concepts
5. 5Mastery model updated from every interaction

Requirements that bind

• Never reveal the answer before diagnosing the gap
• Explanations pitched at the learner's demonstrated level
• Mastery estimates exposed to the learner, not hidden

Stack

Ships with

Dataset

12,000 synthetic learner interactions across 30 topics

Deployment

Docker Compose, Postgres-backed session state

Part of

Architecture

1. 1Disruption event ingested from the operations feed
2. 2MCP servers wrapping inventory, fare-rules and loyalty systems
3. 3Option generation constrained by fare rules and entitlement
4. 4Preference-weighted ranking against the passenger profile
5. 5Autonomous rebooking below a value and confidence threshold
6. 6Warm handoff with full context above it

Requirements that bind

• Never book outside fare rules or entitlement
• Hard spend cap per passenger per disruption
• Idempotent booking — a retry cannot double-book
• Complete decision trail for a later fare audit

Stack

Ships with

Dataset

Synthetic operations feed, 3 fare-rule sets, 5,000 passenger profiles

Deployment

Kubernetes with Temporal workflows and a Redis lock service

Part of

Architecture

1. 1Alarm correlation across topology and time windows
2. 2Agent queries topology, change log and metrics
3. 3Root-cause hypotheses ranked with supporting evidence
4. 4Remediation proposed, never executed unapproved
5. 5Customer notice drafted from the confirmed incident only

Requirements that bind

• No remediation without an on-call engineer's approval
• Hypotheses ranked with evidence, never asserted as fact
• Customer-facing text drafted only after confirmation
• Full trajectory retained for post-incident review

Stack

Ships with

Dataset

Synthetic alarm streams from a 400-node regional topology

Deployment

Kubernetes with Prometheus and an approval webhook

Part of

Architecture

1. 1Portfolio snapshot from the book-of-record
2. 2Market retrieval scoped to held instruments only
3. 3Briefing drafted against a compliance-approved template
4. 4Automated compliance check before release
5. 5Advisor edits captured as training signal

Requirements that bind

• Never generate investment advice — briefing material only
• Every market claim carries a source and timestamp
• Compliance check runs before an advisor ever sees the draft

Stack

Ships with

Dataset

500 synthetic portfolios and 90 days of synthetic market commentary

Deployment

Azure Container Apps with Azure OpenAI and managed identity

Part of

Architecture

1. 1Manual ingestion with section-aware chunking
2. 2Retrieval scoped to the store's own region and format
3. 3Grounded answering with section citation
4. 4Exception filing when no confident answer exists

Requirements that bind

• Answer only from the manual — no general knowledge fallback
• Cite the section number in every answer
• File an exception rather than guessing

Stack

Ships with

Dataset

A 300-page synthetic operations manual across 4 store formats

Deployment

Single Docker image, SQLite + pgvector option

Part of