SciencesLoop Open to AI Engineer roles

Agentic predictive maintenance · deterministic pipeline agent

Chat with a condition-monitoring agent

Type one sentence (try “Scan the fleet”) and a deterministic pipeline agent confirms the plan, runs a 10-stage analysis in front of you (ingest → clean → model → select → predict → diagnose), and hands back a short queue of evidence-backed, cited decision cards — the agent does the heavy lifting, a human makes the calls.

🔒 Deterministic · Grounded · No LLM

The 68-second demo

One sentence → confirmed plan → the agent runs all 10 stages → a decision card gets answered → a cited diagnosis. Try it yourself below.

Try it live

After the free container sleeps, the first visit self-bootstraps in ~2–3 min (downloads the public NASA dataset + trains on the spot). If you see “First boot”, give it a moment — it's instant afterwards. Open full screen ↗

Try:

  • Click “Scan fleet” → confirm the plan → watch the run → answer a decision card
  • Type “diagnose unit 81” → get a bilingual diagnosis with knowledge-base citations
  • Switch to Teaching mode → step through each pipeline stage manually

What it does

Remaining-useful-life prediction, risk banding, and evidence extraction for 100 turbofan engines (public NASA C-MAPSS simulation data), plus a TF-IDF-retrieved maintenance knowledge base that grounds a cited diagnostic report — fully reproducible and auditable (every tool call lands in an execution trace).

Boundaries

Independent R&D prototype: public simulated data, not field telematics; deterministic rule-based orchestration with no LLM in the default path; every output is advisory — triage and sign-off decisions never auto-pass and always require a human. Not affiliated with any equipment manufacturer.

How it works

A rule-based supervisor drives the 10-stage pipeline: each stage runs EXECUTE → VALIDATE gate → pass / raise a decision card / halt. Artifacts carry provenance stamps (checkpoint/resume), the step journal streams into the chat, and every decision-card signal re-derives from a named artifact field. Source and evaluation reports are on GitHub.