AI × Biotech / Bay Area
I build the AI systems a molecular-farming company runs on, and I study how to check the coding agents that build them.
01Now
AI & Automation Lead, since May 2026
“Molecular farming: flavor, fragrance and functional ingredients.”
How Nexus describes itself
I am the architect on the company’s seed-to-sale platform, covering genetics, cultivation, refinement, the QA lab, quality and sales, and I run its internal AI tooling. In biology the data underneath has to be right before any model on top of it can be, so that is where most of the work starts.
PhyTree
I migrated a decade of legacy genetics records (2013 to 2022) into the rebuilt platform. Identifiers disagreed across the old stores, so a pattern-detection pass rebuilt the lineage with confidence scores and gave researchers about 300 parent-child pairs to confirm instead of about 30,000 raw records. PhyTree renders the result as interactive phylogeny trees.
Data Atlas
Every source system lands in cached, auto-refreshing record sets that agents and dashboards query through one MCP server and search API. A provenance graph means any answer traces back to its source record.
246controlled documents made fully searchable
Quality
Release approvals run as a transactional outbox with event-driven fan-out, so a vendor outage becomes a retry instead of a half-completed product release.
Internal platform
Agent schedulers in isolated git worktrees, browser automation and deploy-triage bots, automating work across lab, quality, operations, finance and commercial teams. For the analytical lab that includes GC-FID run planning.
Lunch & Learn
I ran the program department by department, wrote 12 project-knowledge briefs, and set up recurring agent tasks so shared context outlives any single chat.
02Research
Field note, v0.1, in preparation
Field note with data
Rules written after failures: operating multi-agent coding agents at a small company
Most of my engineering runs through coding agents and their subagents. This note records what went wrong, the rules I wrote after each failure, and the layer that ended up enforcing each rule. The analysis code, aggregate data, taxonomy and generic versions of the controls ship with it.
43 active days of retained transcripts, mostly mid-September to October 17
23 of 30
failure classes could be caught by a deterministic check. 13 of those 23 still had no hook, deny rule or code gate.
81 of 86
“failed” entries from a ticket-fixing bot were the harness exiting, not the agent deciding. Read naively, the log overstates agent failure by more than an order of magnitude.
Green ≠ done
A recurring class: a green build, an exit code 0 or a self-written test reported as done, while the behavior was never checked where users are served.
Extract every “done”, “verified” or “deployed” claim and test it against the agent’s own trajectory and against later outcomes, then trace subagent claims through the orchestrator to the human.
Compare rule-based hooks with model-based monitors on the same traffic, at an alert budget one operator can actually absorb.
Measure calibration, omissions and answer flips under pushback when a bot reports status to people who cannot check the work themselves.
Which rules survive a compaction, reach a subagent, or get quietly weakened by a later memory note.
03AI × Bio
In progress
Genetics
Predicting how terpene chemotypes are inherited, using the reconstructed strain lineage as the backbone. Working toward a bioRxiv preprint.
Analytical
ML-assisted peak identification for gas chromatography, anchored to an in-house library of terpene standards.
CommunityA member of the Longevity Biotech Fellowship for over a year, and active in Bay Area lab-automation groups.
04Path
2026 to now
AI & Automation Lead
Seed-to-sale platform architecture, the company data layer, quality systems and internal AI tooling.
2024 to 2026
AI Engineer and Software Engineering Consultant
Agentic QA that drove end-to-end payment flows in place of manual testing; PHI detection at 95%+ confidence with locally hosted models for a HIPAA healthcare platform; voice agents handling hundreds of regulated calls a day for a financial firm; a Kichwa-Spanish translator for indigenous-language education.
2025
B.S. Biotechnology & Computer Science
2023 to 2024
Automation Engineer
A TensorFlow and OpenCV model that estimates plant weight from camera images and started a new toxicology assay; imaging stations wired into Ignition SCADA; a robotic laser-cutter fix for high-throughput screening, projected to recover $450,000 a year in lost runs.
2022
Bioinformatics Research Intern
Metagenomic analysis of DNA samples for an algae-based water filtration system.