Work

Things I built, and where they stop.

Each item says what problem it takes on, what I did, what you can inspect, and what it does not prove. Synthetic scenarios are labelled as synthetic. None of this is paid client work unless it says so.

Start here

Featured

Self-built · not client work biotech-rag-assistant · v1

Biotech RAG assistant: controls before the model

Problem: A controlled-document assistant that ranks by relevance alone will answer from a retired SOP when it is the closest match. And in my own workspace, AI agents produced 107 fabricated citations because a rule I wrote gave them no honest way to say "found nothing".
What I did: I designed and built the assistant trust layer first: validation, a status gate before ranking, a relevance gate that refuses, exact-span citations, and an ablation test that removes the status gate to show what breaks.
Evidence you can inspect: Live demo with deep links to the trap questions, public code, tests, the decision record, and a known gap pinned as a test.
Limits: Synthetic 10-document corpus. Deterministic retrieval: no language model writes answers yet. Not validated, not in production.

Read the case study

Public synthetic REG-CHANGE-001 · v0.1.0

Regulatory-change workflow map

Problem: Regulatory change gets handled as a reading task and a task list. When someone later asks which version of a rule governed the work on a given date, the record cannot answer.
What I did: I wrote the scenario and the map: every step from published amendment to cutover gets an authoritative record, a failure signal, an owner, and the evidence required.
Evidence you can inspect: The full five-column map, plus the method note that explains each column.
Limits: Invented company and labelling scenario. Not a compliance assessment of any real organization, and not legal or regulatory advice.

Open the map

Walkthroughs and collaborations

Public synthetic WER-DEV-001 · v0.1.0

Deviation write-up, reconstructed before automation

Problem: A team wants an assistant to draft deviation narratives. The real investigation already lives across four informal places.
What I did: I mapped the intended and actual paths, separated authoritative records from working copies, and drew the automation boundary.
Evidence you can inspect: Full walkthrough with the path map and boundary.
Limits: Invented QC scenario. Not a client engagement.

Open the walkthrough

Public synthetic WER-QU-001 · v0.1.0

Specification drafting into quality unit approval

Problem: An assistant drafts Rev 4.0 of a specification. One limit has no source, and nobody wrote down what the assistant may invent.
What I did: I traced the evidence chain from draft to approval and showed what document control alone would miss.
Evidence you can inspect: Full walkthrough, anchored to FDA 21 CFR 211.22(c).
Limits: Invented QC scenario.

Open the walkthrough

Permissioned collaboration COLAB-HIRE-001 · v0.2.0

Evidence review on an AI-assisted hiring workflow

Problem: Can a later reviewer recover the sources, versions and human decisions behind an AI-assisted hiring output, or only a clean summary?
What I did: Peer technical collaboration: I walked the workflow with the builder and wrote up the method-level gaps.
Evidence you can inspect: Collaboration note, published with written permission, plus a public LinkedIn recommendation with relationship context.
Limits: Peer collaboration, not a paid client engagement. Configuration, credentials and commercial terms omitted.

Read the note

Public synthetic BRE-SAMPLE-001

Claims audit package

Problem: A product claim set mixes tested capability with validation and compliance wording the evidence does not support.
What I did: I built the full sample package: disposition, claim register, buyer risks, remediation and limits.
Evidence you can inspect: The complete synthetic report on one page.
Limits: No real company, product, client or measured outcome.

Open the package

Older walkthrough WER-OOS-001 · v0.2.0

OOS investigation and an obsolete SOP

Problem: An out-of-specification investigation ran against a superseded procedure.
What I did: My first full method walkthrough: reconstruction questions, evidence map, gap register.
Evidence you can inspect: Full walkthrough.
Limits: Invented scenario. Existing QA would usually catch this failure, so it is kept for method completeness.

View the walkthrough

Code

More on GitHub

Repositories, tests and decision records for the projects above, plus earlier evaluation tools. A failed evaluator stays public as a limit, not hidden.

GitHub profile

Project work

Have a workflow you want looked at?

I also take bounded projects: reconstructing one controlled workflow before it gets automated or moved into a new system. The scope and limits are on the services pages.

Project services

Want to talk about any of this?

Hiring, a project, or a question about how something was built. Pick a reason on the contact form.

Hiring or employment Project or contract