A full-stack ML experimentation and observability platform powered by an Agentic AI engineering co-pilot.
Ask any question about your experiment. An autonomous AI agent — built on Gemini's native function calling — plans a multi-step investigation, executes read-only tools against live telemetry, and returns a structured report with evidence provenance. Grounded in actual experiment telemetry. No hardcoded logic. A real agent loop. ↓ Jump to full details
Ask a question about any experiment. The agent autonomously plans, calls tools, evaluates results, and delivers a structured report — without hallucinating.
The Experiment Investigator is an autonomous AI agent built with Google Gemini's native function calling. It is not a chatbot. It is a bounded, read-only reasoning loop that decides which tools to call, in what order, and when to stop.
User Query
│
▼
┌─────────────────────────────────────────────────────────┐
│ InvestigatorAgent │
│ │
│ ┌──────────┐ ┌───────────────┐ ┌───────────┐ │
│ │ Gemini │────▶│ Tool Dispatch│────▶│ DB / SHAP │ │
│ │ Flash │◀────│ (5 tools) │◀────│ Layer │ │
│ └──────────┘ └───────────────┘ └───────────┘ │
│ │ │
│ ≤ 5 iterations (hard cap) │
│ │ │
│ ▼ │
│ ┌──────────────────────────────────────────────────┐ │
│ │ Structured Report: Conclusion · Evidence · Limits│ │
│ └──────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────┘
5 read-only tools the agent can call:
| Tool | What it fetches |
|---|---|
get_experiment_runs |
All runs in the experiment with status |
get_run_metrics |
Accuracy, F1, precision, recall, ROC-AUC |
get_run_config |
Hyperparameters and model configuration |
compare_runs |
Side-by-side delta between any two runs |
get_feature_importance |
SHAP values and top contributing features |
Built-in safety guarantees: Hard iteration cap (≤5 loops) · Read-only tool access · SHA-256 response caching · 5-model Flash cascade failover · Evidence provenance on every claim.
MicroFlow
├── ML Experimentation Platform
│ ├── Dataset Management & Versioning
│ ├── Experiment Tracking (runs, configs, artifacts)
│ ├── SHAP Explainability (summary, dependence, values)
│ └── Paginated Artifact Registry
├── AI Engineering Suite (Google Gemini)
│ ├── Experiment Investigator ← Agentic, function-calling loop
│ ├── AI Run Review ← Peer-review per completed run
│ ├── AI Run Comparison ← Delta analysis across two runs
│ ├── AI Strategy Co-Pilot ← Chronological experiment guidance
│ ├── AI Dataset Insights ← Schema audit and quality scoring
│ └── Ask MicroFlow ← Hybrid RAG assistant + RAGAS eval
└── Infrastructure
├── FastAPI backend · React/TypeScript frontend
├── PostgreSQL + pgvector (semantic search)
├── Docker Compose (3-container stack)
└── 289+ tests · CI/CD via GitHub Actions
┌──────────────────────────────────────────────────────────┐
│ React Frontend │
│ TanStack Query · TypeScript · Recharts · Tailwind CSS │
└──────────────────────┬───────────────────────────────────┘
│ HTTP / REST
┌──────────────────────▼───────────────────────────────────┐
│ FastAPI Backend │
│ │
│ Routers → Services → Repositories │
│ │
│ ┌─────────────┐ ┌──────────────┐ ┌────────────────┐ │
│ │ Training │ │ AI Layer │ │ RAG Layer │ │
│ │ Engine │ │ (Gemini) │ │ (pgvector) │ │
│ │ + SHAP │ │ + Agent │ │ + RAGAS eval │ │
│ └─────────────┘ └──────────────┘ └────────────────┘ │
└──────────────────────┬───────────────────────────────────┘
│ SQLAlchemy ORM
┌──────────────────────▼───────────────────────────────────┐
│ PostgreSQL 16 + pgvector │
│ Experiments · Runs · Artifacts · AI Cache · Embeddings │
└──────────────────────────────────────────────────────────┘
Key engineering decisions:
- Repository Pattern — No SQL in services. No DB access in the AI layer.
- Zero-Hallucination AI — Gemini receives pre-fetched structured data only. No SQL generation. No internet access.
- Deterministic Caching — SHA-256 hash of every prompt. Identical queries skip the LLM entirely.
- Resilience Cascade — Exponential backoff + 5-model Flash failover on rate limits.
git clone https://github.com/Mohitingale13/microflow-ml-platform.git
cd microflow-ml-platform
# add GEMINI_API_KEY to .env
docker compose up --build| Service | URL |
|---|---|
| Frontend | http://localhost:3000 |
| Backend API | http://localhost:8000 |
| Swagger Docs | http://localhost:8000/docs |
Required .env variables:
GEMINI_API_KEY=your_google_gemini_api_key_here
POSTGRES_USER=microflow
POSTGRES_PASSWORD=microflow_secret
POSTGRES_DB=microflow
All AI features require
GEMINI_API_KEY. Without it, AI endpoints return a descriptive 503 while all other platform features remain fully operational.
docker compose exec backend pytest -v289+ tests covering services, routers, training, AI layer, and the full Investigator agent spec.
microflow-ml-platform/
├── backend/
│ ├── app/
│ │ ├── ai/
│ │ │ ├── investigator_agent.py # Agentic loop
│ │ │ ├── tools/investigator_tools.py # 5 read-only tools
│ │ │ ├── gemini_service.py # Retry + failover
│ │ │ └── cache_service.py # SHA-256 caching
│ │ ├── routers/ # HTTP layer only
│ │ ├── services/ # Business logic
│ │ ├── repositories/ # DB access
│ │ └── training/ # Model Factory + SHAP
│ └── tests/ # 289+ tests
├── frontend/
│ └── src/
│ ├── components/experiments/
│ │ ├── InvestigatorModal.tsx
│ │ ├── AIStrategyTab.tsx
│ │ └── CompareRunsDialog.tsx
│ └── components/datasets/
│ └── AIInsightsTab.tsx
├── docker-compose.yml
└── CONTRIBUTING.md
- Authentication — JWT + role-based access control
- Async training jobs — Celery/ARQ task queue with real-time status
- Cloud artifact storage — S3-compatible backend (single service swap)
- Streaming AI responses — Token-level streaming for the Investigator
- Experiment scheduling — Grid search / random search across hyperparameter spaces
See CONTRIBUTING.md for architecture conventions, layer rules, and PR guidelines.
MIT
Inspired by MLflow, Weights & Biases, and Kubeflow. AI powered by Google Gemini.