Unbiased Neutral Convention for Agnostic Seed Engineering
Privacy-first synthetic data pipeline for fine-tuning LLMs in regulated industries.
Blockchain-anchored quality verification on Polygon PoS.
This project is proudly
Organizations in healthcare, finance, legal, automotive, manufacturing, and education need specialized LLMs — but fine-tuning requires real conversations full of PII, PHI, and legally privileged data.
UNCASE solves this. It generates high-quality synthetic training data from structured seeds (conversation blueprints that contain zero real data), evaluates output against 9 hard-gated quality metrics, and anchors every evaluation on Polygon PoS so auditors can independently verify results without trusting anyone.
Real Conversations → PII Scan → Parse & Validate → 9-Metric Quality Gate
→ Synthetic Generation → Re-evaluate → Blockchain Anchor → LoRA/QLoRA Fine-tune
One pipeline. Zero PII. On-chain proof.
git clone https://github.com/uncase-ai/uncase.git && cd uncase
cp .env.example .env
# Edit .env → set at least one LLM key (ANTHROPIC_API_KEY, GEMINI_API_KEY, etc.)
docker-compose up -d # API + PostgreSQL + Redis + Dashboard
docker-compose ps # Verify all containers are running- API: http://localhost:8000 (Swagger at
/docs) - Dashboard: http://localhost:3000
Optional profiles:
docker-compose --profile ml up -d # + MLflow (port 5000)
docker-compose --profile gpu up -d # + NVIDIA GPU support (port 8001)
docker-compose --profile observability up -d # + Prometheus (9090) + Grafana (3001)Requires Python 3.11+ and PostgreSQL 16+.
git clone https://github.com/uncase-ai/uncase.git && cd uncase
uv sync --extra dev
cp .env.example .env # Set DATABASE_URL + at least one LLM key
make migrate # Run database migrations
make api # Start API on port 8000
# Optional: start the dashboard
cd frontend && npm install && npm run devpip install uncase # Core
pip install "uncase[ml]" # + transformers, peft, trl, torch
pip install "uncase[privacy]" # + SpaCy, Presidio NER
pip install "uncase[blockchain]" # + web3, Polygon PoS anchoring
pip install "uncase[all]" # Everything| Feature | What it does | |
|---|---|---|
| Privacy | Privacy Interceptor + PromptShield | Zero PII tolerance — Presidio NER, SpaCy, 9 regex patterns, adversarial scan |
| Quality | 9 hard-gated metrics | ROUGE-L, Factual Fidelity, TTR, Coherence, Semantic Fidelity, Embedding Drift, Tool Validity, Privacy, Memorization |
| Verification | Blockchain anchoring | Every evaluation SHA-256 hashed → Merkle tree → Polygon PoS. Verifiable via Polygonscan |
| Generation | LLM Gateway | Route to Claude, GPT-4, Gemini, Groq, Ollama, vLLM — privacy-intercepted |
| Import | Connector Hub | WhatsApp exports, webhooks, HuggingFace, CRM, custom connectors |
| Export | 10+ formats | ChatML, Llama, Mistral, Qwen, Nemotron, Alpaca, ShareGPT — all with tool-use |
| Tools | 30 domain tools | 5 per industry across 6 verticals, with simulation and training data |
| Sandbox | E2B cloud | Parallel generation in isolated MicroVMs, instant demo containers |
| SDK | Python SDK | Sync + async programmatic access via httpx |
| MCP | MCP server | Expose tools to Claude Code and any MCP-compatible agent |
| Plugins | Marketplace | 6 official plugins (one per industry), extensible architecture |
| Enterprise | Auth + observability | JWT, audit logging, cost tracking, rate limiting, Prometheus + Grafana |
| Domain | Namespace | Tools | Plugin |
|---|---|---|---|
| Automotive Sales | automotive.sales |
5 | automotive |
| Medical Consultation | medical.consultation |
5 | medical |
| Legal Advisory | legal.advisory |
5 | legal |
| Financial Advisory | finance.advisory |
5 | finance |
| Industrial Support | industrial.support |
5 | industrial |
| Education Tutoring | education.tutoring |
5 | education |
Each includes specialized seed templates, quality thresholds, compliance rules, and built-in tools.
| Layer | Purpose |
|---|---|
| 0 — Privacy | PII detection + PromptShield adversarial scan (audit/warn/block) |
| 1 — Parser | Multi-format parsing (CSV, JSONL), auto-detect OpenAI/ShareGPT/UNCASE formats |
| 2 — Evaluator | 9 quality metrics, composite scoring, blockchain anchoring |
| 3 — Generator | LLM-powered parallel generation with tool-augmented conversations |
| 4 — LoRA Pipeline | LoRA/QLoRA fine-tuning with DP-SGD (epsilon ≤ 8.0) |
106+ API endpoints across 25 routers. 1,590+ tests at 73% coverage.
See docs/architecture.md for the full system diagram and database schema.
| Document | Description |
|---|---|
| Architecture | System diagram, SCSF layers, database schema |
| API Reference | Complete endpoint map with examples |
| Features | Detailed feature docs with usage examples |
| Configuration | Environment variables, Docker services, pip extras |
| Development | Setup, testing, CLI, contributing guide |
git checkout -b feat/my-feature
make check # lint + typecheck + tests
uv run pytest tests/privacy/ # mandatory before PRSee docs/development.md for the full guide.
BSD 3-Clause — Free for commercial and non-commercial use.
UNCASE — Because the best training data is data that never existed.