Chem AI

A local, private, single-user tool for market researchers scoping new factories or business expansions in the chemical, pharmaceutical, and semiconductor industries. You give it a rough idea; six specialist departments — Market Analysis, Procedure, Procurement, Intellectual Property, Finance, and Presentation — debate their way to a full research pack.

The flow

1. You submit a rough factory / expansion idea. One or two paragraphs; upload any relevant context PDFs or Word docs (existing feasibility notes, patent search reports, vendor quotes).

2. Aarav (the Analyst) reads it and asks the smallest set of clarifying questions needed — product, industry, target scale, geography, budget ceiling, timeline, regulatory posture, existing capabilities. You answer inline. Loop until the idea is sharp enough — Aarav tracks a completeness score. If the idea is vague, Aarav will keep asking until it's not.

3. You lock the idea. Aarav writes a structured Refined Concept.

4. Six specialist departments debate their artifacts one at a time, with you as the gatekeeper between every department. Each department writes independent initial drafts, then in each subsequent round critiques teammates' drafts and revises its own, self-scoring agreement 0-100. Debate stops as soon as every member's score hits the threshold (default 95%) or the max round cap (default 4). The department lead's final draft becomes the artifact of record, at which point the pipeline halts and shows you an Approve / Revise panel:

   • Approve → the next department starts immediately.
   • Revise (with feedback) → the same department re-runs from round 1 with your feedback baked into every prompt. You can revise as many times as needed; each cycle is logged.

   Departments can also halt themselves mid-debate: if a member hits a genuine ambiguity in round 1, the department raises a Clarification panel with the specific questions it needs answered. Blank answers are allowed — the department will document them as assumptions. Once you submit, the debate resumes from round 2 with your answers as ground truth.

   • Market Analysis (3) — Ananya (lead) + Karthik + Meera → B2B Market Analysis: TAM/SAM/SOM, buyer segments, top 5 geos, competitors, pricing reference, Go / Conditional / Not-Recommended verdict.
   • Procedure (5) — Dr. Prakash (lead) + Sneha + Ravi + Aditi + Dr. Nikhil → Procedure & Route of Synthesis: chosen route, reaction scheme (Mermaid), step-by-step conditions (T / P / moles / yield), material balance, lab-scale + industrial-scale configuration, organic- chemistry mechanism / stereochemistry / impurity review, hazards & EHS register, waste/effluent plan.
   • Procurement (3) — Vikram (lead) + Rohit + Priya → Procurement Plan: Equipment BOM + Raw Material BOM (India + international vendors, landed cost analysis, single-source flags).
   • Intellectual Property (2) — Arjun (lead) + Kavya → IP Analysis: Product / process / intermediate patent landscape (India + US + EU + WIPO), FTO verdict, design-around suggestions, own filing strategy.
   • Finance (3) — Neha (lead) + Suresh + Divya → Financial Projection: CAPEX + 5-yr P&L + cash flow, per-unit manufacturing cost, sales projection, IRR, sensitivity analysis (INR ₹ primary + USD reference).
   • Presentation (2) — Ishaan (lead) + Tara → Presentation Package: Board-ready executive summary + full detailed report + appendix with consolidated Open Questions.

5. Read the pack. Each artifact is a proper Markdown document with the right sections, requirement IDs (MKT-, PROC-STEP-, PROC-HW-, PROC-RM-, IP-, FIN-), and traceability. Drill down into the debate rounds for critique and per-member drafts.

Pipeline (sequential, per-department Approve gate)

Only one department runs at a time. After every department finishes its debate you are prompted to Approve or Revise before the next department starts — the pipeline never proceeds without an explicit "yes" from you.

Refined Concept (Aarav — asks clarifying Qs if vague, loops until ready)
    │
    ▼
  1. Market Analysis                    ← Approve to continue
    │
    ▼
  2. Procedure OR Semiconductor Mfg     ← Approve to continue
     (one runs per session, chosen by session.industry)
    │
    ▼
  3. Procurement  (uses Procedure + Market)      ← Approve to continue
    │
    ▼
  4. Intellectual Property  (uses Procedure + Market)   ← Approve to continue
    │
    ▼
  5. Finance    (uses Procurement + Market)      ← Approve to continue
    │
    ▼
  6. Presentation   (aggregates everything)      ← Approve to complete

Each department may also raise mid-debate Clarification questions after round 1; the pipeline pauses there too until you answer.

Stop / Retry (recovery controls)

• Stop pipeline (header button, visible while a run is in flight): cancels the current run. The in-flight LLM call finishes in the background (the Cursor SDK can't interrupt a call mid-flight), but the session flips to cancelled immediately and the department that was running is recorded as Cancelled by user.
• Retry stage (button on any errored / cancelled tile): re-runs just that department. Prior revision history and clarification Q&A are preserved so context is not lost. Once the retried stage finishes, the pipeline continues from there.
• Start over (only visible after a cancel): wipes every artifact and re-runs from department 1.

Together these cover the "network glitch mid-run" recovery path: if a stage fails (or you Stop and want to resume), click Retry on the red tile and the pipeline picks up where it left off.

If the server itself was killed mid-run (crash, stop.sh, deploy), the next server start runs a one-shot sweep (sweepInterruptedSessions) before it accepts traffic. Any session that was persisted as generating is a zombie by definition — no orchestrator is running its pipeline anymore. The sweep marks the interrupted stage as errored with the message Server restarted before this step completed. Click Retry to resume. and flips the session to error. Open the session and click Retry on the highlighted tile — approved upstream stages, revisions, and clarification Q&A are all preserved. The only degradation is uploaded document context (the parsed docs live in an in-memory cache and don't survive restart); the refined concept and every prior stage are on disk, so the retry still runs.

Every specialist is fully configurable — name, model, tone, role description, model parameters, avatar, accent — and persisted to your browser's localStorage. Add or remove department members freely, as long as the per-department min/max are respected (min 2, max 4 per department; Procedure is capped at 5 to accommodate a dedicated organic- chemistry expert).

Defaults

• India-first: INR ₹ (with USD reference), Indian regulatory regime (BIS, PESO, CPCB, DCGI-CDSCO), IP jurisdictions include India + US + EU + WIPO. Any of these can be overridden by the Analyst per session.
• Generic Procedure department: the same 5 specialists adapt to chemical / pharma / semiconductor from the Refined Concept — they don't need a per-industry configuration.
• B2B only: Market Analysis is strictly B2B (formulators, OEMs, foundries, distributors) — no consumer / retail analysis.

Requirements

• Node.js 22 or newer
• A Cursor API key (see .env.example)

Run it

./start.sh

• Web UI: http://localhost:5277
• API: http://localhost:5278/api/health

Stop with Ctrl+C or ./stop.sh.

Coexisting with architecture-group/: this app runs on ports 5277 (web) / 5278 (api), while architecture-group/ uses 5275 / 5276 — you can run both at once.

UI tabs

• New Idea — start a session.
• Refine — interactive Q&A loop with Aarav.
• Pipeline — live view of the currently-running department, the Approve/Revise panel when a department is waiting on you, and the Clarification panel when a department raised mid-debate questions. The full DAG is always visible so you can see what's approved, what's running, and what's queued.
• Documents — final artifacts, one tab per document, plus a collapsible drill-down into every debate round.
• My Team — configure departments (2-4 members each; Procedure 2-5) and the debate settings (agreement threshold + max rounds).
• History — every past session with its full artifact pack and how each department terminated (converged vs. max rounds).

Data

• History and session state → server/.data/
• Specialist config + debate settings → browser localStorage (keys mr.specialists.v1, mr.settings.v1)
• Logs → .logs/

Nothing leaves your machine except LLM calls, which go directly to the Cursor SDK.

Requirement-ID conventions

Downstream departments trace claims back to these IDs. If a claim has no source, it's flagged as an assumption or open question rather than invented.

Department	ID prefixes
Market Analysis	MKT-FIND-###, MKT-ASSUMP-###, MKT-OQ-###
Procedure	PROC-STEP-###, HAZ-###
Procurement	PROC-HW-###, PROC-RM-###
IP	IP-PROD-###, IP-PROC-###, IP-INT-###, IP-USE-###, FTO-###
Finance	FIN-RM-###