I build the evidence layer that makes agent work checkable.
Proof-Driven Development — the method, in one page. No framework. No install.
Proof-Driven Development rests on three rules:
- Default FAIL. Every criterion starts
false. The agent cannot mark it passing without opening the evidence first. - The proof must be able to fail. If the behavior broke tomorrow, would this check turn red? If not, it is decoration.
- The judge never built it. The grader gets a fresh context and no write tools.
Rule 2 is the one most systems skip. It is also the one that finds fake evidence.
| Context | What I built | Result |
|---|---|---|
| Meta Product Quality & Experience (embedded via Tests Assured) | Planted-bug evaluation for an agentic QA platform | 1.00 precision, 0.80 recall, 0.89 F1; rerun cost −98% |
| JPMorgan Commercial Banking | Agentic-RAG diligence over 100,000+ documents, human review preserved | Prospecting 4 hr to 15–30 min; a 2,000-company analysis 3 weeks to under 15 min |
| AWS DeepRacer, JPMorgan Palo Alto team | Reward-function and race-line optimization | Lap time 23.260 s to 18.121 s (theoretical line optimum: 19.673 s) |
Rule 2 is not theory. Planted bugs are how I learned that a passing check often proves nothing.
| Layer | Repo |
|---|---|
| Doctrine | proof-driven-development |
| Reference implementation | NodeProof — executable proof gates that block unsupported completion claims |
| Components | NodeRL (task evaluation) · NodeTrace (surface to span) · NodeMem (memory gates) · agentic-ui-qa (artifact-only completion) · AgentRedteam |
| Instruments | FeatureClipStudio (walkthrough clips as reproducible proof media) · NodeAgent (the distilled core, with its public conformance suite) |
| Applications | NodeRoom — the flagship: a live human-agent workspace, in production at noderoom.live · NodeBenchAI (entity intelligence, live at nodebenchai.com) · NodeVoice (shared-state voice rooms) |
One idea. The rest are parts of it. If you only open two: the doctrine page, then NodeRoom — the method and the thing it built.
All 19 projects, grouped by layer →
3.5 years inside JPMorgan commercial banking: credit underwriting, covenant analysis, and startup banking. 72 credit transactions, roughly $800M in aggregate exposure, 270 financial models. Then I became the engineer who builds the AI that automates that work.
The role this points at is forward-deployed / applied AI engineering: strongest where regulated, document-heavy workflows need agents somebody can actually check.
LinkedIn · nodebenchai.com · hshum2018@gmail.com
Feedback wanted. Rule 2 does not scale by hand. Breaking each behavior on purpose is cheap for three checks and expensive for three hundred. If you have solved this for agent workflows, I want to hear how.