Reenvisioning quality control for manufacturing operations

Unplanned downtime drains an estimated $1.4 trillion annually from global manufacturing giants. To address these operational losses, ETH Zürich spinoff Forgis (a graduate of the Google DeepMind Robotics Accelerator) is building the world model for industrial systems, a foundation modelHyperion, trained on machine telemetry that understands how a line behaves, why it fails, and what it will do next. ItThis foundation model turns that understanding into root cause analysis, event prediction, and autonomous robotic intervention across the automotive, food & beverage, industrial equipment, and machine building sectors.

By pairing its proprietary industrial world model, Hyperion, with customized Gemma models, Forgis can deploy intelligent quality control systems directly onto factory edge hardware to predict mechanical failures in advance, isolate underlying root causes, and guide floor operators through corrective steps before defects occur.

Factories using Forgis’ system have seen 60% reduced downtime, 30%, improved efficiency, and 70% less scrap compared to pre-deployment.

Solving the edge bottleneck for high-stakes production

Factory floors demand real-time insights and continuous reliability without network dependencies, so the team needed a reasoning engine that could run locally while sharing a single edge GPU card with Hyperion, its multimodal world model.

After benchmarking several open-weights models, including Qwen 2.5 3B, Llama 3.2 3B, and Phi-3.5-mini, the Forgis team chose Gemma 3n E4B. The model’s Per-Layer Embeddings (PLE) offload embedding tables from the edge GPU into host RAM, while layer-wise KV cache sharing minimizes time-to-first-token latency for floor operators.

Data sovereignty is enforced by the architecture rather than by policy. Hyperion compresses each machine’s live condition into a 256-dimensional embedding, and that compression is irreversible: raw signals cannot be reconstructed from it. Nothing else is ever passed to the language layer, which cuts bandwidth by more than a thousandfold and leaves the rest of the edge GPU's budget free for generation.

For a lot of our customers, the cloud has simply been off the table: privacy rules and real-time latency keep the data on-site. With Hyperion reading the machine and a fine-tuned Gemma explaining it, all on the edge, they finally get a foundation-model-grade answer right at the machine, with nothing leaving the plant.

Riccardo Maggioni

CTO, Forgis

In manufacturing, seconds dictate costly errors; running locally, the system can detect anomalies in real time. For labeling machines, Forgis’ platform anticipates registration drift before parts breach reject limits, recommending a temporary web-tension adjustment and scheduling a roller replacement for the next changeover, turning what would be an unintentional scrap run into a two-minute planned fix.

Forgis' HMI helps flood operators monitor issues. — Forgis’ FABRICA platform on HMIs helps operators monitor and preempt issues in manufacturing production lines.

Grounding language models in physical machine states

Standard language models cannot natively parse raw, multi-sensor industrial time series. Given a real vibration signal alongside an incorrect fault label, frontier models confidently explain the wrong answer. Hyperion overrides the label and diagnoses from the signal itself.

The same embedding that protects the data carries the diagnosis. Hyperion builds it by fusing live sensor streams with static priors held in each machine`s documentation and maintenance logs, so it captures not just what the machine is doing now, but where it is heading and why. The model does this by predicting representations rather than raw values, focusing on degradation trends and fault signatures instead of sensor noise. That objective needs no labels, so the representation is learned from ordinary production data. Only a small event-prediction head on top is trained on labels, and because the representation already carries the signal, it reaches 92% of full-supervision accuracy on lifecycle tasks with 2% of the annotations.

A tokenizer then discretizes the state encoding that Hyperion produces into a short sequence of machine tokens drawn from a specialized 125-token vocabulary. Forgis adds these 125 machine tokens directly to Gemma 3n’s vocabulary and customizes the model for each machine by fine-tuning it on its specific manuals and maintenance logs. This enables Gemma to translate live machine states into immediate root-cause explanations and concrete fixes for human operators.

In die-casting applications, when Hyperion translates recent production shots into machine tokens, Gemma recognizes emerging porosity patterns. Grounded in the technical documentation, it isolates the physical causes, such as an early fast-shot changeover entraining air, delayed or low intensification pressure leaving trapped gas uncompressed, or elevated melt temperatures increasing dissolved hydrogen, and prescribes the exact corrective adjustments before scrap limits are exceeded.

Hyperion builds representations for multiple downstream tasks. It enables classification and prediction of both seen and unseen faults, also cross-machine diagnoses. A dedicated Gemma, fine tuned per machine, reads the representations and answers to the operator grounded in what the sensors are doing right now, not a guess from the manual.

Jonas Petersen

Head of Research, Forgis

Giving an industrial world model a voice

Forgis empowers operators to optimize performance by harnessing the immense volume of operational data generated daily across factory floors.

Deploying Gemma alongside Hyperion directly at the machine proves that advanced foundation models can thrive within the physical, air-gapped constraints of the factory edge. By converting multivariate sensor streams into actionable, natural-language guidance, Forgis enables operators to act before failures and defects occur and brings general-purpose machine intelligence to the factory floor.