What am I actually qualified to do? 20 yrs Systems biologist / biotech founder, 3 years of daily agentic-AI building, no CS background. other advanced degrees. Pull no punches.
retained text
Asking for calibration, not selling anything. Systems biologist, 20+ years, PhD-track training, between jobs. I founded a biotech that shipped hardware, software, consumables, SaaS and edge compute. It works, but an exit is years off. For three years, 4 to 8 hours a day, I have built with ChatGPT, then Claude Code, local MLX models for housekeeping and other agents. I write the specs, invariants and gates. Agents write all the code. I run my business, a budding consultancy and primary healthcare research on it. None of it is paid work yet. \*\*The principles I built on\*\* (each came from a failure, not a blog): Build systems with a closed self-improving feedback loop. The system will then if managed get smarter than you quickly at its tasks. 1. \*\*Make the wrong state unconstructable.\*\* Derive, never declare. Closed registries, hash-derived ids, trust class computed from the source type. If an author can type it, an author can fake it. 2. \*\*A deterministic gate sits downstream of every non-deterministic model.\*\* Pure-Python checks decide what ships; the LLM only proposes. 3. \*\*Well-formed is not true.\*\* Every claim carries a verbatim quote, and a "literal floor" rejects any claim whose numbers, dates or units are not inside that quote. 4. \*\*Cite the authority, not the transport.\*\* Search engines and tool servers locate evidence; the primary source is the citation. 5. \*\*Anything an agent can set is not a gate.\*\* Irreversible actions need a human at a TTY typing a phrase. No flag, env var or config can open it. 6. \*\*Incident to invariant to test.\*\* An agent deleted a delivered email; that became a spec, a plan and 8 test-first commits. 7. \*\*The working tree is scratch; the commit is storage.\*\* Every write carries a millisecond timestamp and links, in one atomic command. \*\*The system those principles produced:\*\* \- \*\*Operating contract.\*\* One \~8,000-word [CLAUDE.md](http://CLAUDE.md) with hard gates, reversibility tiers, and \~17 hand-written hooks that enforce the ones that can be enforced. \- \*\*Trust layer.\*\* A fact engine with the ontology written before the code: \~3,000 one-line atoms, closed predicate and origin registries, a reject-only local LLM judge keyed to each atom's hash. Hand-rolled, not RDF or SHACL. \- \*\*Tool layer.\*\* A fork of ToolUniverse (\~3,000 biomedical tools: FDA, ClinicalTrials, ChEMBL, UniProt, SEC, USPTO) with \~120 return schemas hardened and \~1,100 contract examples; a Google Workspace CLI; a "Live API" mode where the agent gets the dev docs, one guidance file and the auth core, backed by a catalog of 97 systems and 18,291 endpoints. \- \*\*Pipelines.\*\* Biotech due diligence: dataroom in, investor report out, every digit checked against a cached source quote, adversarial review by DeepSeek and Kimi before publish. \~50 reports, 20+ companies, 100+ runs. Prompt optimizer with bootstrap CIs; it recorded a tie, not a win. \- \*\*Automation.\*\* Launchd jobs in one TOML registry with a two-way drift test. A fact-lock on the LinkedIn publish step. \- \*\*Scale.\*\* \~130k lines of Python in seven repos, as many lines of tests, CI green, 100% branch coverage on the gated subset. \*\*What is missing:\*\* nothing containerized, nothing runs off my Mac, "observability" is JSONL ledgers, eval samples too small to separate arms, 20 upstream PRs all closed unmerged, no code I typed myself. Single-host internal tooling. Not an MVP, not production. \*\*Questions:\*\* 1. In October 2026 terms, what level is this? 2. Which gaps block a scientific-lead forward-deployed or applied-AI role at a life-science company? Which are noise? 3. If you have hired someone like this, what evidence changed your mind? 4. For a domain expert who steers agents rather than types: how do you answer "what did you personally build"? Original repos are private. The ToolUniverse fork is public. Tell me what you would need to see.
document fetched 2026-10-21 00:39 · canonical https://reddit.com/r/ClaudeCode/comments/1xavpha/what_am_i_actually_qualified_to_do_20_yrs_systems