The 33 artifacts behind the register, each with its kind, the claim it supports and a short summary. Each title links to the artifact itself.
Deployed a grounded RAG chatbot that passed a 33/33 golden-QA gate before ship.
Website rebuilt + chatbot answering only from the business's real content; eval suite re-run against the live URL post-deploy.
Hand-built the first instance end to end; the reusable template and its seven skills were extracted from it.
Hand-built grounded-chat architecture, independent prospect corpus and brand.
Built a public demo where anyone can watch a real traced pipeline run step through its stages, then run the actual dashboard themselves via a one-command Docker bundle.
Live product case-study site with an animated real-trace pipeline stepper and a runnable Docker demo.
Clone → docker compose up → working analysis dashboard in ~5 minutes.
Public compose bundle; pipeline IP excluded via verified sanitized images.
Every input ID survives high-volume LLM batch runs; failed rows are automatically re-run rather than silently dropped.
Every input ID accounted for; failed rows re-run at batch-size-1 with strict-JSON coaching.
Integrated Bedrock via langchain_aws so structured outputs pass schema validation on arrival, and built a throttling ladder tuned to Bedrock's actual per-minute quota instead of a generic backoff curve.
Three fast retries then a deliberate 61s wait past the quota boundary, jittered; typed and string-wrapped ThrottlingException both caught; timeouts handled on a separate shorter path. Part of an AWS migration incl. Bedrock evaluation and GPU self-hosting of Llama 70B.
Requests fall back through an ordered provider chain on failure, retry against schema validation, and every call's cost is logged automatically.
Durable retry-payload queue on exhaustion; free-model-only governance for autonomous surfaces.
Workflows pause mid-stage for human approval and resume from the gate.
LangGraph-equivalent constructs implemented from scratch (see equivalence diagram).
Keyword and vector retrieval fused by RRF, degrading gracefully rather than failing, and running in the live demos today.
Sibling server-side implementation uses Postgres FTS + pgvector hybrid search.
Built and shipped a hand-written MCP server (5 tools, 3 resources, 3 prompts) with auth and per-tool rate limits enforced.
From a consumer AI product in private beta (capabilities disclosed, concept not).
Chatbots must pass the full golden suite before and after deploy.
must/mustNot assertions incl. leak guards; sample fictional-client case set included.
Picked the shipped model against explicit pass/fail gates on quality, cost, and latency, with the report to show the reasoning.
Golden corpus, 5-dimension rubric, p50/p95 tracking, CSV export.
Derived per-job cost from token mechanics; caught a 6-7x pricing error in code constants.
Standard vs batch/flex tables across six models, with reconciliation addendum.
Authored IaC for a production-shaped product; separately extended an upstream AWS stack and used selected AWS services for development and performance work.
State/vars excluded; AWS (ECS/Batch/SQS) stack credited as upstream, extended by me.
Built and exercised a deployment pattern with isolated per-engagement stacks, automatic TLS and per-engagement secrets.
Includes a documented Go reverse-proxy duplicate-header fix.
Proves proprietary code is absent from every published image layer.
docker save + per-blob marker scan; avoids the rm-in-a-later-layer trap.
Generates tailored documents that never invent a fact, with keyword-coverage checks and a human checkpoint before anything ships.
Never-invent rule, two-skin design system, ATS-safe rendering, full test suite.
Won't generate outreach copy for a claim it can't back up with proof.
Lead-facts schema, offer/sector crosswalk, visual-regression + smoke suites.
Built the consultation-analysis application solo: pipeline, backend, review frontend, infrastructure and docs; delivered it for a paying client, then took it into a 2026 government consultation of more than 2,000 comments on self-hosted open-weight models inside New Zealand.
Configurable multi-stage pipeline, hierarchical themes, audit-trailed human review, NZISM-aware RFP work, and self-hosted open-weight inference run on New Zealand research compute.
Productised the platform pattern; on a public academic benchmark (SemEval-2014), scored in-house, topic detection comes to 0.83 F1 in the product's discovery mode. The client-engagement validation is the stronger evidence, and the case study ranks them.
Discovery-mode detection 0.83 F1 (product-as-sold, no gold categories seeded); predetermined-mode polarity roughly level with the best 2014 system (0.84 vs 0.829, like-for-like); demo ships separately.
Turns a prospect's URL into a live grounded chatbot and a priced proposal, through 7 human-gated stages.
Intake → scrape → design → build → chatbot evals → ship → proposal; two live instances.
Built a provider-agnostic, file-system-first agent platform with the same shape as an organisational agent deployment, and run my own daily work through it.
Skills registry, staged workflows with human gates, retrieval with golden tests, free-tier model routing with an append-only cost ledger.
Mapped my orchestration engine to LangGraph primitives, concept for concept, so the equivalence can be checked directly.
nodes/state/checkpoints/tools ↔ stages/run-state/pause-gates/skills.
Mapped AWS services used in development work, plus upstream AWS foundations extended, to Azure equivalents with provenance stated per service.
Bedrock↔Azure OpenAI/Foundry and EC2↔NC-series derive from development/performance work; five further mappings are upstream foundations extended; Azure remains architecture-only.
Built the n8n automation layer joining enquiry, CRM, calls, follow-up, enrolment records and Student Services handover across a distributed admissions operation.
Selective AI for language and transcript work inside a rules-driven customer journey; enquiry replies moved from 24+ hours to about 10 and enrolment handover from two working days to the same day; the nine workflows are rebuilt on synthetic data as Exhibit 20.
Extracted and sanitized the legislation RAG built for Stoa: section-level bill corpus, hybrid retrieval, citations a reviewer can verify.
Extracted from the consultation-analysis application: legislation parsed into a section-level corpus, hybrid retrieval over it, and theme-to-section mappings grounded in verifiable citations.
Segmented ~800 recorded sales calls on conversational shape and content, then tested the segments against real conversion outcomes.
TF-IDF reduced before it meets hand-built conversation features; silhouette search over feature subsets and k. Ships with a synthetic fixture; the real transcript corpus is customer PII and is excluded.
Found that tuned ensembles beat a linear baseline by nothing, and shipped the linear model with learning curves as the evidence.
Linear vs tuned Random Forest vs XGBoost in one pipeline, learning-curve overfitting diagnosis, and a one-standard-error rule that declines to prefer a lead smaller than its own measurement noise.
When the model can't place a response, the pipeline degrades how specific the answer is rather than inventing one: a fallback ladder that climbs the theme hierarchy and then stops, recording the response as unmapped instead of fabricating a finding.
From Stoa's unattended analysis runs over tens of thousands of public submissions. Leaf mapping, then a lenient re-prompt, then progressive fallback up the parent hierarchy, then a deliberate refusal to force a mapping. Ships alongside the provider-level retryability classifier that separates transient failures from terminal ones.
Built the portfolio voice agent first on OpenAI Realtime/WebRTC, then migrated the live implementation to Gemini Live/WebSockets, retaining streamed audio, interruption handling, transcription and short-lived session credentials.
Versioned transport evidence: earlier OpenAI WebRTC implementation; current Gemini Live WebSocket implementation; provider keys remain server-issued and short-lived.
Built an agent-action boundary where models may draft but cannot execute: actions require visible user confirmation, signed authorization, strict validation and idempotent server-side recording.
Draft-only tool contracts, a separate confirmed-action envelope, short-lived signed authorization, durable two-action session limits and failure handling that records notification status without duplicating the external effect.
Built isolated per-recipient AI context using hashed bearer exchange, signed HTTP-only sessions, allowlisted browser projections, integrity validation, expiry and revocation.
The raw bearer is never stored; only its SHA-256 hash reaches D1. Every session restore revalidates packet integrity, expiry and revocation, while private notes and source material remain server-side.
Rebuilt the nine admissions workflows as n8n files that import and run without credentials: every service and model call carries pinned synthetic output, and every rule runs for real.
Rule nodes check model output against the course sheet before a person sees it, human gates hold the run until an adviser or Admissions responds, and one script imports all nine into a fresh n8n, runs every branch and compares each rule's output with its expected file.