A production-grade platform for building, deploying, and monitoring ML/AI applications.
Atlas AI gives teams the full stack — from model training and deployment to real-time observability, drift detection, and incident management. Built around real use cases (network intrusion detection, account takeover, RAG) but designed to be model-agnostic.
🔗 Live UI: https://vigilant-ui.duckdns.org
🔗 Live API: https://vigilant-api.duckdns.org
🔗 LinkedIn post: https://www.linkedin.com/posts/bara-alsedih_mlops-machinelearning-python-activity-7458189356126863360-7yBI
Tech: Python · FastAPI · PostgreSQL · ClickHouse · Polars · scikit-learn · React · TypeScript · Vite · Docker · Caddy · Poetry
| Repo | Purpose | Stack |
|---|---|---|
| vigilant-api | Monitoring backend — evaluation reports, drift detection, incident lifecycle, telemetry | Python · FastAPI · PostgreSQL · ClickHouse |
| vigilant-ui | Observability dashboard — overview, evaluation, drift, incidents | React 18 · TypeScript · Vite · Tailwind · TanStack Query |
| vigilant-detect | ML service — ATO and network intrusion detection with automated weekly retraining | Python · FastAPI · scikit-learn · Polars · APScheduler |
| vigilant-rag | RAG application — document retrieval and LLM-powered responses over local models | Python · FastAPI · Ollama |
| vigilant-pack | CLI runtime — boots any AI application stack (services, models, app) in one command | Python · Click · Rich · Docker |
Each repo owns its own setup, env vars, tests, and tag-triggered deploy workflow.
- OLTP (PostgreSQL) — model registry, evaluation reports, incidents. Transactional, mutable, relational.
- OLAP (ClickHouse) — production traffic log, alerts, drift results, report metrics, LLM traces. Append-only, high-volume.
| Prefix | Purpose |
|---|---|
/api/v1/reports |
Evaluation report history (latest, last 10) |
/api/v1/reporter |
Pre/post-production evaluation runners (data, model, drift) |
/api/v1/monitoring |
Live drift & telemetry metrics |
/api/v1/incidents |
Incident lifecycle (auto-triggered by alerting engine) |
/api/v1/telemetry |
System health & latency probes |
Full reference: vigilant-api README.
Drift Detection — statistical shifts between training reference and incoming production data via PSI, KS-test, and Chi² test.
Performance Monitoring — accuracy, F1, precision, recall, and confusion-matrix deltas over time. Alerts on decay relative to the pre-production baseline.
System Health — API latency and schema consistency via middleware. Alerts on slow requests (>500ms) and 5xx errors.
Automated remediation is defined in vigilant-api/core/procedures.yaml. Low-risk incidents auto-resolve; high-risk ones create tickets for human review.
| Incident Type | Risk | Behavior |
|---|---|---|
system_latency |
Low | Auto-resolves (refetch DB) |
schema_skew |
Low | Auto-resolves (refresh schema) |
data_drift |
High | Creates incident ticket |
performance_drop |
High | Creates incident ticket |
All services run on a single Oracle Cloud Always-Free VM behind Caddy (automatic Let's Encrypt TLS). Each repo has a tag-triggered GitHub Actions workflow:
deploy.api.*→ ships vigilant-apideploy.ui.*→ ships vigilant-uideploy.ml.*→ ships vigilant-detect
Bara Al-Sedih — github.com/baraalsedih · baraalsedih@gmail.com