GenComply
AI practice

AI engineering, built to reach production

Most AI projects die between the prototype that works and the system that ships. We staff that gap.

Stage 01
Prototype
Works in a notebook. Metrics look good.
The gap · we staff here
Latency · perimeter · behaviour · evidence · drift
quantisation on-prem inference bounded actions eval harness retraining
Stage 02
Production
Deployed, monitored, auditable, still running in month twelve.

Where projects stall

None of these are model quality problems. They are engineering problems, and they are what our AI engineers spend their time on.

01The model performs in a notebook and cannot meet the latency budget in the real system.
02The data cannot leave the perimeter and the architecture assumed a hosted API.
03The agent behaves in testing and does something unspecified in week three.
04Nobody can explain a specific output to the person asking.
05There is no retraining pipeline, so the model quietly degrades.

Four practice areas

01

Custom and edge models

PyTorch · Hugging Face · PEFT/LoRA · ONNX Runtime · TensorRT · OpenVINO · llama.cpp · NVIDIA Jetson · MLflow

Model selection, fine-tuning on proprietary data, quantisation and distillation, deployment to embedded targets, industrial PCs and on-premise GPU. No inference leaving your infrastructure.

02

LLM agents and workflow automation

Python · LangGraph · vLLM · FastAPI · pgvector · Qdrant · Redis · OpenTelemetry

RAG pipelines on enterprise data, agents with explicitly bounded action sets, human-in-the-loop approval paths, complete decision logging. Predictable behaviour, which matters more than clever behaviour when the agent touches a regulated process.

03

Anomaly detection and predictive maintenance

scikit-learn · PyTorch · Evidently · InfluxDB · TimescaleDB · Grafana · MQTT · OPC UA · Modbus

Detection on real equipment signal, tuned against your false-positive tolerance rather than a public benchmark. Deployed at the edge, with drift monitoring.

04

AI governance and assurance

Model inventory · Eval harness in CI · Decision records · Limitation docs · Assurance pack

Model inventories, evaluation harnesses in CI, decision records, and evidence packs assembled for the supervisor you actually answer to.

How we staff AI work

Same three engagement models. ML engineers into your existing team, a dedicated AI squad, or full delivery of a defined system.

Most AI engagements start with a scoping phase, because AI scope written before the data has been examined is fiction.

Contact

Tell us what you need staffed.

Send the role, the stack, the engagement model and the start date. We reply within two business days with whether we can cover it and who would be on it.

Reply time2 business days
Working hours09:00–18:00 CET
Not billedScoping calls, interviews
No sales sequence. One reply from an engineer.