Foundation models, fused into one understanding.

01

Time-series forecasting

TimesFM 2.5 is the workhorse for zero-shot equipment-health forecasting (vibration RMS, temperature, current draw), OEE projection, throughput, and yield. We ensemble with Chronos for quantile bands and Moirai for irregular sampling intervals. Temporal Fusion Transformer is the fallback when operators need per-feature attribution alongside the point forecast.

02

Vision-Language-Action (robotics)

OpenVLA (UC Berkeley, Apache 2.0) — first commercially-usable open-weight VLA, trained on the Open X-Embodiment dataset of 1M+ trajectories — drives cobots and AMR manipulators. Octo (93M params) covers fast-inference edge embodiments. RT-2 and π₀ remain the closed-weight benchmarks we measure against. Cross-embodiment transfer means a policy trained on a UR5 transfers to an ABB GoFa with only a small fine-tune.

03

Machine vision

Qwen3-VL handles complex defect taxonomies via prompt — adapt to a new client standard by editing instructions, not retraining a CNN. Pixtral runs at the cabinet edge in 25 W envelopes. Both deliver structured findings — defect class, severity, location — that flow into the work-order system without a human in the loop unless severity exceeds a threshold. Battery-cell deployments using this pattern run 100% inspection at 0.3 sec/cell with documented 97%+ first-pass yield.

04

Acoustic anomaly detection

PANNs CNN14 embeddings with per-asset linear classifiers. Each transformer, gearbox, motor, or pump is enrolled once at commissioning (under 30 min of audio). The platform watches each asset's acoustic-emission signature continuously and flags deviation: bearing wear, gear-tooth fault, stator/rotor imbalance, fan unbalance. AE catches faults earlier than vibration alone thanks to higher frequency sensitivity.

05

Reinforcement-learning process control

Soft Actor-Critic and Proximal Policy Optimization for continuous-process loops (refining, blending, fermentation, paper machine wet-end). Control-Informed RL embeds PID set-point tracking inside the deep-RL architecture so the controller gains nonlinear modelling capacity without surrendering engineering trust. Validated on published 800,000 t/a catalytic-reforming benchmarks; deployed inside the customer's safety envelope at all times.

06

Reasoning & embeddings

Gemini 3.1 Flash Lite as the conversational agent and tool-orchestrator. Tool calls are typed (zod-validated) and per-tenant. Gemini Embedding 2 in Matryoshka 1536-d for the workspace tier — every site has a long-term memory of events, anomalies, work orders, deviation reports, and agent decisions, semantically searchable, retrieved by the agent before acting.

07

Provenance, end-to-end

Every prediction is stored with model name, version, weights hash, input window, and confidence interval. Operators can replay a forecast a year later with the same inputs and verify the result is identical. For autonomy this matters: every dispatch carries the full lineage of how its prediction was produced.