Cameron Sanderson

Process before platform.

I build and test AI-assisted systems, bringing a regulated-industry approach to evidence, quality and how work actually gets done.

About ten years in pharmaceutical QC and sterile compounding labs. The software and AI work is self-directed, and each project says what it does not prove.

Cameron Sanderson

Selected work

Two projects worth opening first

Each one names the problem, what I did, what you can inspect, and where it stops.

Self-built · not client work

Biotech RAG assistant: controls before the model

Problem
A document assistant will happily answer from a retired SOP if it is the closest match.
What I did
I built the trust layer first: a status gate before ranking, a refusal when nothing approved supports the answer, exact-span citations, and tests that remove each control to show it matters.
Evidence
Live demo, public code and tests, and the decision record.
Limits
Synthetic 10-document corpus. Deterministic: no language model writes answers yet. Not validated, not in production.
Read the case study
Synthetic scenario

Regulatory-change workflow map

Problem
Regulatory change usually gets tracked as a task list. Months later, nobody can show which version of the rule governed the work on a given date.
What I did
A five-column record chain for every step (step, authoritative record, failure signal, owner, evidence required), worked through on an invented labelling change.
Evidence
The full map, plus the method note behind it.
Limits
Invented company and scenario. Not a compliance assessment of anyone, and not regulatory advice.
Open the map

The method that transfers

What a regulated lab taught me about AI systems

Quality work is mostly about whether a record will still hold up when someone asks later. The same question applies to AI output.

Process before platform

Work out how the job actually runs before encoding it. Most of the automation problems I see started as workflow problems.

Check the evidence, not the confidence

An answer with a citation attached is not the same as a supported answer. I build tests for the difference.

Give every check an honest way to fail

If a rule leaves no compliant way to say "I looked and found nothing", something will fake compliance. 107 fabricated citations taught me that one.

See the case study

State the limits

Every project here says what it does not prove. A working demo is not production readiness.

Background

From sterile compounding to software

I've spent about ten years in pharmaceutical quality control and sterile compounding labs, most recently in cleanroom aseptic compounding. That's where the habits come from: records that have to hold up later, and a healthy suspicion of anything that looks finished.

I now build and test AI-assisted systems, and I'm looking for work where both halves are useful.

Hiring, a project, or a question about something I wrote?

Send a short note and pick a reason, so it lands in the right place. Or write to [email protected].