Prooflane is an evidence-first AI venture studio. It turns a startup idea into a decision-ready workspace containing market research, a business model, financial projections, risk analysis, validation experiments, an AI mentor, and investor exports.
The product is intentionally more than a text generator. It keeps assumptions, evidence, experiments, generation progress, contradictions, and deterministic financial checks as structured records that founders can inspect and update.
- Eleven-stage LangGraph workflow for research, strategy, finance, risk, validation, cross-section auditing, and final compilation.
- Live web research with inline sources and refreshable evidence claims.
- Evidence-adjusted viability scoring that discounts unverified conclusions.
- Editable assumptions, evidence reviews, and validation experiments.
- Persisted background generation with progress, cancellation, resume, and startup recovery.
- Deterministic financial consistency checks for funding allocations and unit economics.
- Authenticated plan library and plan-grounded AI mentor.
- Decision, risk, market, business model, financial, and full-plan views.
- Generated PDF report, PowerPoint pitch deck, and Excel financial workbook.
- Production-mode Docker images for Next.js, FastAPI, and PostgreSQL.
Requirements: Docker Engine or Docker Desktop with the Compose plugin.
cp .env.example .env.localAdd at least one valid Gemini API key to .env.local, replace SECRET_KEY with
a long random value, then run:
./setup.shOpen:
- Product: http://localhost:3000
- API: http://localhost:8000
- Interactive API docs: http://localhost:8000/docs
Useful commands:
make ps
make logs
make health
make check
make smoke
docker compose downmake smoke-full performs a real, end-to-end AI generation and consumes external
API quota.
Browser / Next.js
|
| JWT + JSON / generated files
v
FastAPI API ───── PostgreSQL
|
v
Persisted generation job
|
v
LangGraph supervisor
market → identity → strategy → capabilities → finance
→ deterministic finance check → risk → growth → validation
→ contradiction audit → compiler
|
+── Gemini structured generation
+── DDGS live research
+── ReportLab / python-pptx / OpenPyXL exports
Generation returns 202 Accepted immediately. Work progresses between durable
stage checkpoints, and the frontend polls the generation job rather than holding
one fragile request open.
.
├── backend/
│ ├── app/
│ │ ├── api/ # Authentication and plan HTTP endpoints
│ │ ├── models/ # SQLAlchemy persistence models
│ │ ├── nodes/ # Specialist workflow agents and auditors
│ │ ├── schemas/ # API and structured-LLM contracts
│ │ ├── services/ # Workflow, research, jobs, scoring, exports
│ │ └── utils/ # Authentication and document builders
│ └── scripts/ # Deterministic and integration smoke checks
├── frontend/
│ ├── app/ # Next.js routes and product workspace
│ ├── components/ # Report, chart, progress, validation UI
│ └── lib/ # API client
├── docs/
│ ├── API.md
│ ├── ARCHITECTURE.md
│ └── OPERATIONS.md
├── docker-compose.yml
├── Makefile
└── setup.sh
The repository uses layered checks:
# No network or database: financial arithmetic and evidence scoring
docker compose exec -T backend python scripts/check_deterministic.py
# No network: build and parse PDF, PowerPoint, and Excel
docker compose exec -T backend python scripts/check_exports.py
# Authenticated database/API CRUD
docker compose exec -T backend python scripts/smoke_validation_workspace.py
# Live research refresh; uses network and API quota
docker compose exec -T backend python scripts/smoke_research_refresh.py
# Full live generation and structured persistence
docker compose exec -T backend python scripts/smoke_full_generation.py
# Production frontend compilation
docker compose exec -T frontend npm run build- AI output is decision support, not verified investment, legal, or financial advice.
- Search results are research leads. A founder must open and verify a source before marking an evidence claim verified.
- Background jobs currently execute inside the FastAPI process. A horizontally scaled deployment should move them to a dedicated queue/worker.
- Database tables are created automatically at startup. Production deployments should add versioned schema migrations before evolving the model.
- Authentication is single-user ownership with JWTs; there is no RBAC, team workspace, password reset, email verification, or server-side token revocation yet.
- The browser stores the access token locally. A production hardening pass should use secure, HTTP-only cookies and add rate limiting.
MIT © 2026 Henok Yoseph.