Context layers that let industrial AI reason — not guess.
MechFrame is an R&D effort to build LLM-ready context layers for industrial plants: control loops, tag semantics, process topology, and rollups — so a model can answer engineering questions with traceability instead of hand-wavy “AI vibes”.
What we’re building
Most “AI for manufacturing” fails because models don’t know what tags mean, how equipment connects, which signals belong to a loop, or what “normal” looks like. MechFrame’s bet is that the right context layer structure makes an LLM behave more like a junior process engineer: curious, grounded, and falsifiable.
Semantic tag catalog
Units, ranges, descriptions, aliases, and “what this signal actually represents” — so the model stops mixing up flow, level, and valve position.
Control-loop awareness
SP/PV/CV relationships, mode history (auto/manual), limits, and loop groupings — enabling loop-centric questions instead of raw tag hunting.
Process topology graph
Nodes/edges that encode equipment connectivity and influence paths — so “upstream/downstream” and cause-effect reasoning becomes possible.
Rollups for fast answers
Precomputed windows (like 1h/12h/24h) for quick comparisons and “what changed” queries without grinding through full-resolution history.
Roadmap
The near-term goal is a trial portal where a small number of plants can safely test the system. Login is a placeholder today — designed so we can later plug in real auth + a chat UI.
Phase 1 — Trial portal
Static login template → authenticated chat UI → per-user sessions and audit logs.
Phase 2 — Grounded responses
Every answer links back to tags/loops/rollups used, with “why” and “what to check next”.
Phase 3 — Plant playbooks
Reusable templates for common areas (stock prep, paper machine, evaporators, recaust) to speed up onboarding.
Phase 4 — Alerts + investigations
Detect drift, instability, and manual operation patterns, then auto-build an investigation trail.