- 01
Microcontroller
64 KB – 2 MB, no NPU
- 02
Embedded Linux
512 MB – 2 GB
- 03
Edge accelerator
up to 67 TOPS
- 04
On-prem server
GPU
- 05
Cloud
elastic
Each practice names the range of compute it works in. Most engagements touch two or three of them.
Six practices, one continuum
The interesting engineering is not in the model. It is in making a model that was trained on a workstation survive on a part with kilobytes of RAM, no accelerator, and a power budget it has to share with everything else on the board.
The layer that decides whether any of the rest ships. Bring-up, drivers, capture pipelines and the unglamorous work of making a build reproducible on someone else's hardware.
Three sensing modalities, one discipline: turning a raw signal into a decision close to where it was captured. Radar is the one most teams skip, and often the one the application actually needed.
Most generative work fails on retrieval, not generation. We build the corpus handling and evaluation first, then decide how small the model can be, including small enough to run with no connection at all.
The classical work, done properly. Method chosen to fit the problem class and the data you actually have, rather than one approach applied to everything.
Before a model is worth building, the data has to arrive reliably and mean the same thing every time. Often the hardest part is the system that already exists and cannot be replaced.
Capabilities
- Post-training quantization and QAT
- Pruning and architecture reduction
- Cycle and memory budgeting
- Power profiling and duty-cycle design
- On-device inference runtimes
- Accuracy recovery after compression
Tools
TensorFlow Lite Micro · ONNX Runtime · CMSIS-NN · DeepCraft · Hailo · Jetson · PSoC6 · SAMA7 · STM32MP
Capabilities
- Board bring-up and BSP work
- Embedded Linux, Yocto and Buildroot
- MIPI camera and sensor pipelines
- Video capture and streaming
- Python to C++ conversion and optimisation
- Cross-compilation and release builds
Tools
Yocto · Buildroot · GStreamer · V4L2 · OpenCV · C/C++ · Rust · Zephyr · FreeRTOS
Capabilities
- Object detection and tracking
- Surface and PCB defect inspection
- Face detection, recognition and liveness
- FMCW radar: presence, counting, gesture
- Range-Doppler and angle-of-arrival processing
- Keyword spotting and acoustic anomaly detection
Tools
60 GHz FMCW · raw ADC capture · CFAR · YOLO family · RetinaFace · OpenCV · multi-vendor radar front-ends
Capabilities
- Retrieval over manuals and technical corpora
- Chunking, embedding and hybrid retrieval
- On-device LLMs and offline assistants
- Vision-language models for anomaly description
- Agent and tool-calling pipelines
- Evaluation sets and regression testing
Tools
RAG · hybrid retrieval · rerankers · quantized 1B–8B models · VLMs · streaming TTS · tool-calling agents
Capabilities
- Predictive maintenance and failure modelling
- Forecasting and demand modelling
- Dimensionality reduction on wide feature sets
- Structured experimental data analysis
- Anomaly and outlier detection
- Model evaluation, drift and monitoring
Tools
scikit-learn · PyTorch · XGBoost · time-series methods · dimensionality reduction · experiment tracking
Capabilities
- Pipeline design and orchestration
- Warehouse and lakehouse modelling
- Legacy and informatics system integration
- Data quality, lineage and validation
- Streaming and batch ingestion
- Platform migration without downtime
Tools
Python · SQL · Airflow · dbt · Spark · Kafka · Snowflake · Databricks · Foundry · on-prem warehouses