🧠 Company Brain Workbench

Full-stack AI employee memory layer

When a human resolves an ambiguity once, the system turns that judgment into reusable company memory. The next AI employee does not ask again.

---

The bottleneck in AI employees is not task execution.
It is judgment reuse — turning one human decision into a durable rule the whole organization can apply.

Company Brain Workbench is a full-stack product prototype for human-in-the-loop AI operations: ambiguous cases route to a human once, the correction becomes structured memory, and future matching work resolves automatically.

⚡ The Demo in 60 Seconds

1	Priya Sharma joins as a Sales Engineer in Dubai. The workbench extracts the case, detects that “Sales Engineer” can map to Sales or Engineering, and stops instead of guessing.
2	A human resolves the ambiguity: Sales Engineer = Sales. The decision creates a Company Brain rule with downstream impact: CRM access, sales-core group, and sales-field finance tier.
3	Omar Reyes joins as a Sales Engineer in Dubai. The Company Brain rule fires automatically. Omar resolves to Sales with no human escalation.

Corrected once. Never asks again.

🖼 Product Screens

Empty brain	Rule learned and applied

🚀 Run

Backend

cd backend
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
PYTHONPATH=. uvicorn app.main:app --reload --host 127.0.0.1 --port 8000

Backend docs:

http://127.0.0.1:8000/docs

Frontend

In a second terminal:

cd frontend
npm install
NEXT_PUBLIC_API_BASE=http://127.0.0.1:8000 npm run dev

Open:

http://localhost:3000

🏗 Architecture

├── backend/
│   ├── app/main.py                    FastAPI app · local CORS · API routes
│   ├── app/company_brain/
│   │   ├── engine.py                  Intake, ambiguity detection, rule matching
│   │   ├── schemas.py                 Pydantic request/state models
│   │   └── store.py                   Atomic JSON state persistence
│   ├── tests/test_company_brain.py    Priya → rule → Omar regression test
│   └── data/.gitkeep                  Runtime state directory · JSON ignored
│
├── frontend/
│   ├── app/page.tsx                   Product shell and guided demo UI
│   ├── app/styles.css                 Restrained internal-tool design system
│   └── lib/api.ts                     Typed frontend API client
│
├── assets/
│   ├── company-brain-workbench-demo.gif
│   ├── company-brain-workbench-initial.png
│   └── company-brain-workbench-completed.png
│
└── .github/workflows/ci.yml           Backend tests + frontend build

🎯 Design Decisions

Decision	Why
Ambiguity is a first-class product state	The system should not hide uncertainty behind confident language. It should expose the decision and its downstream impact.
Human correction becomes structured memory	A resolution is not just an answer; it becomes a reusable rule with pattern, decision, source case, and application count.
The hero flow is deterministic	Priya always escalates first; Omar only auto-resolves after the rule exists. The demo is reliable for walkthroughs and recording.
Audit trail is visible by default	Enterprise trust comes from proving what happened: case received, ambiguity detected, approval created, rule created, rule matched.
One accent color marks learning	The interface stays calm; blue is reserved for the moments where company memory is created or applied.

🔐 Security Posture

This is a local product prototype, not a production deployment.

Safeguard	Status
.env files ignored	✅
Runtime JSON state ignored	✅
Local CORS origins only	✅
Input length validation	✅
Unsupported control-character rejection	✅
No raw HTML rendering of case data	✅
Dependency audit clean at push time	✅
GitHub Actions CI	✅

See SECURITY.md for production hardening notes.

🧪 API Surface

POST /api/workflows/intake          Submit an employee/ops case
GET  /api/approvals                 List human-resolution items
POST /api/approvals/{id}/resolve    Turn a human answer into a rule
GET  /api/brain/rules               Inspect learned Company Brain rules
GET  /api/audit                     Review the decision trail
POST /api/demo/reset                Reset local demo state
POST /api/demo/run                  Execute Priya → Omar guided demo

🧬 Related Demo

This workbench expands the original focused thesis demo:

AI Employee Console — a lightweight product demo built around the same “corrected once, never asks again” idea.

💡 The Company Brain Thesis

Every time a human resolves an ambiguity an AI employee could not handle, that judgment should become structured memory: what happened, what context mattered, what was decided, and when it should apply again.

Over time, this layer compounds. Each correction eliminates a future escalation. Each deployment makes the next one smarter.

The moat is not one AI agent — it is the accumulated decision intelligence of the organization.

---

Built by Karuna Sagar

Founder, Plato Tech — clinical & document AI agents shipped to hospitals