📌 Current snapshot — 2026-10-07
Live
Status
- 🟢 ML Capstone — Course 2 (CRISP-DM): Done
- 🟢 ML Capstone — Course 3 (Human Factors): Done
- 🟢 Technical PoC — Phase 1 (prototype): Done
- 🔵 Technical PoC — Phase 2: In progress — the M1–M6 milestones below are its scope.
Milestones — Phase 2
- ✅ M1 — Dataset & Model — XGBoost model trained offline on the AI4I 2020 predictive-maintenance dataset. The demo fleet is scored by the model (pre-calculated predictions with SHAP drivers, risk direction and motor useful life), no longer simulated.
- ✅ M2 — CLAUDE.md & Specs — agentic Spec-Driven workflow (OpenSpec) established.
- ✅ M3 — AWS Deployment — EC2 + HTTPS, plus a GitHub → AWS CI/CD pipeline (push to
main auto-deploys via OIDC + SSM, no SSH). Images are built in CI and pulled by production; TLS renewal lives in version control.
- ✅ M4a — Pre-visit briefing (voice out) — on-demand briefing generated with Amazon Bedrock and read aloud via the Web Speech API, with a deterministic fallback.
- ⚪ M4b — Post-visit report (voice in) — not started: record audio → Whisper transcription → pre-fill the report form.
- 🔵 M5 — Observability & Orchestration — delivered on the local stack: OpenTelemetry instrumentation with a Collector exporting to local Grafana and Grafana Cloud, telemetry ingestion with a real daily inference job, and self-hosted n8n orchestration — one distributed trace spans n8n → backend → PostgreSQL. Remaining: run the orchestration tier in production (production still serves the pre-calculated predictions) and natural-language scenarios that drive each lift's daily risk status.
- ⚪ M6 — Maintenance feedback loop — not started: reported visits steer generated telemetry back toward healthy, so flagged lifts can resolve.
Recently shipped
- M5 on the local stack: OpenTelemetry + Grafana, telemetry ingestion with real inference, n8n orchestration.
- Explainability: SHAP direction (risk-driving vs protective factors) and temperatures in °C.