Science
Methodology, cited end-to-end.
We run an explicit model bench. Every prediction the platform makes is traceable to the model, version, dataset, and confidence interval that produced it. Where research models exist, we use them. Where domain physics exists, we use that too. We do not train novel foundation models — we orchestrate the best of what the field has produced.
Model stack
Open foundations, cited and named.
We orchestrate the best open and research-grade models for industrial systems — never black boxes. Every prediction, every dispatch decision is traceable to the model, dataset, and confidence interval that produced it.
Read the methodologyTimesFM 2.5
Google Research
Equipment health + OEE forecasting
Vibration RMS, kurtosis, and crest-factor trend forecasting; OEE / yield / throughput projection at line and station granularity. Zero-shot on new asset classes.
Chronos
Amazon
Probabilistic ensemble partner
Quantile-aware forecasts for failure distributions; remaining-useful-life with calibrated tails.
Moirai
Salesforce
Mixed-frequency forecasting
Heterogeneous SCADA + MES + lab-data fusion at irregular sampling intervals.
Temporal Fusion Transformer
Open research
Interpretable multi-horizon forecasting
Per-feature attribution for production planners — helpful when an OEE drop has multiple plausible drivers.
OpenVLA
UC Berkeley
Vision-Language-Action for robotic manipulation
Open-weight (Apache 2.0) cross-embodiment robot policy trained on the Open X-Embodiment dataset (1M+ trajectories). Drives cobots, pick-and-place, assembly tasks, agnostic to robot vendor.
RT-2
Google DeepMind
Reference VLA (closed weights)
55B-parameter Google DeepMind policy. Strong generalisation; not publicly available, used as a benchmark target for OpenVLA fine-tunes.
Octo
UC Berkeley
Lightweight cross-embodiment policy
93M parameters, fast inference, runs on edge silicon for smaller cobots and AMR manipulators.
Qwen3-VL
Alibaba
Quality-inspection VLM
Defect detection on production-line imagery, OCR on serialised parts, dimension reads, packaging integrity. Adapts to new defect taxonomies via prompt rather than retraining.
Pixtral
Mistral
Edge VLM for in-line inspection
Fits a Hailo-10H / Jetson Orin Nano Super at the line edge — sub-second inspection at 1,200+ parts per minute throughput.
SAM 3.1
Meta AI
Spatial segmentation
Asset masks for cropped, model-ready imagery; product segmentation on conveyor; safety-zone monitoring around cobots.
PANNs
Open research
Acoustic-emission classification
Bearing wear, gearbox tooth fault, motor stator/rotor fault, fan unbalance — continuous, low-power, in-cabinet enrolment from a few minutes of audio per asset.
YAMNet
Edge audio embeddings
Always-on monitoring at machine cabinets — emits embeddings on-device so raw audio stays inside the customer's perimeter.
ProcessSim ensemble
Ipnops · open foundations
Discrete-event + process simulator
Production-line balancing, predictive vs reactive maintenance scenario runs, AGV/AMR fleet routing, batch-recipe optimisation, autonomous-truck dispatch. OpenDESS / SimPy / Aspen-style foundations.
SAC / PPO controllers
Open research
Reinforcement-learning process control
Soft Actor-Critic and Proximal Policy Optimization for continuous-process loops (refinery catalytic reforming, blending, fermentation). Control-Informed RL hybrids preserve PID set-point tracking while gaining nonlinear modelling capacity.
Gemini 3.1 Flash Lite
Google DeepMind
Reasoning + tool orchestration
Conversational interface, tool-calling agent for forecast/simulate/dispatch, document QA over standards (ISA-95, IEC 62443, ISO 10218), shop-floor reports.
Gemini Embedding
Google DeepMind
Semantic search over operational text
SOPs, work instructions, MOC records, deviation reports, batch records. Matryoshka 1536-d for pgvector halfvec at the workspace tier.
Benchmarks
Tested against the published baselines.
OEE benchmark:Nakajima TPM’s world-class OEE benchmark is 85% (90% Availability × 95% Performance × 99% Quality). Real-world US plants typically run 55-60%; world-class discrete is 85%; world-class continuous-process is 90%+. The platform tracks every improvement against this canonical reference.
Inspection vision: Production-line inspection systems in 2026 (Cognex In-Sight 3800-class, VisionPro Deep Learning) operate at 1,200+ parts per minute with documented 99.2%+ defect detection accuracy on automotive body panels, semiconductor wafer lines, PCB assembly, and pharmaceutical tablets.
Battery cell inspection: 2026 gigafactory deployments inspect every cell in 0.3 seconds with 100% coverage and 97%+ first-pass yield, detecting defects down to 20 microns — well below human-hair scale.
VLA robotics:OpenVLA (UC Berkeley, Apache 2.0, trained on the Open X-Embodiment dataset of 1M+ trajectories) matches Google’s closed RT-2 performance on cross-embodiment manipulation. Octo (93M params) covers fast-inference edge embodiments.
Acoustic-emission monitoring: AE sensors detect rotating-machinery faults earlier than vibration-only sensors thanks to their higher frequency sensitivity. PANNs CNN14 pretrained on AudioSet provides strong zero-shot anomaly behaviour; per-asset linear classifiers fingerprint each transformer or gearbox in under 30 minutes of enrolment audio.
Standards
Native to the protocols of the floor.
Industrial systems run on a thicket of standards. A platform that doesn’t speak them natively is doomed to perpetual integration work. We speak them natively.
ISA-95 + Unified Namespace
Enterprise-to-shop-floor data architecture
OPC UA + MQTT Sparkplug B
Edge data movement; ISA-95 over MQTT
IEC 61131-3
PLC programming languages (LD, FBD, ST)
IEC 62443
Industrial automation cybersecurity (zero-trust IACS)
NIS2 (EU 2026)
Mandatory cybersecurity for critical-infra entities
ISO 10218 + ISO 13849
Robotic safety; speed-and-separation monitoring
MTConnect
CNC + machining-center telemetry
ROS 2
Robot Operating System for cobots, AMRs, manipulators
ISO 50001 / 50006
Energy management (industrial)
ISO 55001
Asset management (predictive maintenance grounding)
GMP + Pharma 4.0 (PAT)
Pharmaceutical manufacturing compliance
FSMA 204
Food traceability (24h record retrieval)
Selected references
Where the work comes from.
OpenVLA: An Open-Source Vision-Language-Action Model
Kim et al. — UC Berkeley, 2025 (arXiv:2406.09246)
RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control
Brohan et al. — Google DeepMind, 2023
Open X-Embodiment: Robotic Learning Datasets and RT-X Models
Open X-Embodiment Collaboration, 2023 (1M+ trajectories)
Decoder-only Foundation Model for Time-Series Forecasting
Das et al. — TimesFM, 2024 (rev. 2026)
Recent Advances in Reinforcement Learning for Chemical Process Control
MDPI Processes, 2025
Control-Informed Reinforcement Learning for Chemical Processes
Industrial & Engineering Chemistry Research, 2024
Computer vision in precision livestock farming
PMC Animal Frontiers, 2024
Qwen3-VL: A Multimodal Large Language Model Family
Qwen Team, Alibaba 2026
PANNs: Large-Scale Pretrained Audio Neural Networks
Kong et al., IEEE/ACM TASLP 2020
A Review on Acoustic Emission Technique for Machinery Condition Monitoring
ScienceDirect, Mechanical Systems and Signal Processing 2024
NIS2 Directive — EU Critical Infrastructure Cybersecurity (in force 2023; transposition 2024-2026)
European Commission
IEC 62443 — Industrial Automation and Control Systems Security
IEC, ongoing series