Building NodeRoom — a live room where humans and AI agents do high-trust research together, without clobbering each other.
A career, compiled: banking/finance → data engineering → agentic AI, converging on human-agent collaboration systems where the agent leaves receipts.
Meta · agentic QA (PQX) · JPMorgan · 3.5 yrs, credit + agentic-RAG over 100k+ docs · Ideaflow · Founder, NodeBench AI · UC Santa Barbara · full history ↗
| Repo | Layer | What it is |
|---|---|---|
| noderoom ⭐ | Current flagship | Live multi-panel room where humans + NodeAgents edit a shared spreadsheet, note, and post-it wall through one versioned concurrency model — lock → draft → smart-merge, no-clobber, per-element CAS. |
| nodebench-ai | Research engine | Entity intelligence for any company, market, or question — searches + synthesizes with sources, turns each run into a reusable artifact, and ships a hosted public-research MCP. |
| NodeAgent | Agent kernel | The distilled core of NodeBench — one loop, four tool UIs: live context, grounded/cited search, a versioned spreadsheet delta, and a TipTap notebook memo. |
| feature-walkthrough-gif | Proof / media | Playwright → Remotion → ffmpeg turns any feature into an annotated walkthrough GIF — and because it's scripted, the GIFs double as an integration smoke-test. |
| local-collab-mvp | Room OS | Why co-located voice agents collapse into "yeah, exactly…" loops — and the fix: one server-authoritative room state (floor control, deterministic reducer), proven by side-by-side bad/good traces and a measured 6-model coordinator eval. |
| proofloop | Completion gate | Bring any coding agent — proofloop blocks "done" until an executable proof gate passes: Stop-hook refusal, agent-tamper-proof proof state, fail-closed tool contracts. Zero runtime dependencies. |
Productivity infra: gmail-workspace-public (large inbox → one queue, one decision; private data stays local, public research delegated to NodeBench) · agent-workspace-template (reusable Convex/Next agent-workspace runtime).
Also fresh (Jun–Jul 2026): noderl (agentic-RL substrate: trace → reward → memory → repair; 100/100 benchmark tasks scored with zero answer-key writers) · NodeMem (passive agent memory behind a 6-gate suppression pipeline) · visual-judge (deterministic browser evidence + a Gemini video critic) · solo-founder-agent-builder (one-prompt build run passing 32/33 executable proof gates).
📜 Context: TIMELINE.md — the field's viral agent moments (ELIZA → scratchpads → AutoGPT → Manus), the layer every one of them was missing, and where each repo below sits on that map.
flowchart LR
NB["NodeBench AI<br/>research / diligence engine<br/>sourced dossiers · MCP"]
NA["NodeAgent<br/>distilled agent kernel<br/>one loop · four tool UIs"]
NR["NodeRoom<br/>CURRENT FLAGSHIP<br/>live room · lock→draft→merge"]
RO["Room OS<br/>shared-state voice agents<br/>floor control · loop suppression"]
PF["Proof<br/>walkthrough GIFs + proofloop gates<br/>demos that double as tests"]
NB -->|distill the core| NA
NA -->|put humans + agents in one room| NR
NR -->|many agents, one floor| RO
NR -->|ship review-ready artifacts| PF
style NR fill:#111,color:#fff,stroke:#111
Five capability buckets, each load-bearing in the work above:
- Banking & diligence — 3.5 yrs at JPMorgan: credit analysis (72 deals, ~$800M, 270 models) plus "LLMsuite," an agentic-RAG diligence tool over 100k+ documents. Turning messy research into structured, cited sheets and risk models — the reason NodeRoom is a War Room, not a toy.
- Data engineering — pipelines, schemas, reactive runtimes (Convex), durable streaming. The plumbing under every live room and report.
- Agentic AI & evals — agentic QA at Meta (PQX) and eval pipelines at Ideaflow: grounded search, tool loops, versioned model deltas, LLM-as-a-Judge scoring, scenario-based tests. Agents that get checked, not trusted — the harness matters more than the model.
- Healthcare / regulated workflows — prior-auth auto-fill with validation + eval: structured extraction where being wrong has consequences.
- Product engineering — Next.js / React / TS surfaces, UI parity harnesses, reproducible demos. The artifacts people actually click.
Multiple humans and multiple NodeAgents research companies in one live room and enrich a shared diligence sheet together:
- Agents claim an affected-range lock (still readable as context), a blocked agent drafts around it, and on unlock the draft smart-merges — committed human edits are never clobbered. Every edit carries a per-element version (CAS).
- Findings stream into the sheet, the note panel, and the post-it wall — no refresh; server-led agent work reaches every client (e.g. the live
Q3DEMOroom,/ask reconcile Q3 revenuefilling a variance column). - Runs two modes from the same code: a deterministic no-key in-memory engine + scripted agents (
npm run demo), and Live with a real Convex reactive backend + a model-routed LLM agent (routes promoted by ladder evidence, not provider brand). - Ends with downstream-ready review artifacts: company brief, runway chart, open-questions list.
People + agents + artifacts + evidence + review + shareability.
📚 Selected earlier work — the arc that compiled into the systems above
| Project | Signal |
|---|---|
| parity-studio | Image (or live app route) → verified componentized ui_kit, self-judged on a 16-check deterministic rubric with honest score drift. |
| LLM Prior Authorization | LLMs auto-fill prior-auth forms from patient notes — structured extraction, validation pass, LLM-as-a-Judge eval scored on clinical knowledge. |
| Banking assistant | Finance-document assistant for company/PDF analysis — the diligence reflex, pre-NodeBench. |
| openai-agent-eval-framework | Agent evaluation for classification, context verification, and pruning — the eval discipline, early. |
| CosmaNeura med billing | ICD/CPT recommendation from physician dictation — regulated extraction before the prior-auth system. |
| FluencyMed | Early healthcare AI workflow prototype. |
| voice_email_agent | Email ingestion, summarization, embeddings, voice query — the seed of the Gmail workspace. |
The agent should leave receipts. Sources on every claim, a version on every edit, an eval on every answer, and a demo anyone can reproduce. High-trust work doesn't get faster by trusting the model more — it gets faster by making the model checkable.
Autonomous task horizons are doubling roughly every seven months. When AI labor becomes schedulable, the hard questions stop being "how smart" and become who owns the state, who audits the decisions, who holds the floor, who can pause the fleet. Chat windows don't answer those questions — rooms do. That's the bet behind everything above (the full timeline →).
📫 LinkedIn · hshum2018@gmail.com