Applied AI / ML Engineering
Bespoke machine learning and AI pipelines, firmware and robotics, and data analytics. Any industry, engineered to last.
Custom engineering, end to end: from first prototype to production systems your team can run on its own.
Systems built around large language models: knowledge management, product oracles that know your product cold, chat bots, and content generation. Each ships with the retrieval and evaluation plumbing underneath, so it holds up in production.
LLM Orchestration / RAG / Agents / Evals
Hard as it is to believe, many problems are still best solved with traditional machine learning: no slow, expensive LLM required. Classifiers, forecasters, and anomaly detectors that run in milliseconds on hardware you already own, plus a straight answer when that is all your problem needs.
Classification / Forecasting / Anomaly Detection / Recommendation
Legacy codebases ported to new languages and architectures, with inputs and outputs validated against the original at every step so behavior survives the move. Then the extensions: cleaner interfaces, test coverage where there was none, new capability layered onto software you already rely on.
Incremental Migration / Behavioral Validation / Test Harnesses / API Design
Digging into the shape of your data to find what is buried in it: profiling, exploratory analysis, and the unglamorous cleaning that real answers depend on. When a one-off analysis needs to become routine, it gets a custom transformation pipeline built around it.
Exploratory Analysis / Data Cleaning / Transformation Pipelines / Reporting
From feasibility study to working prototype in weeks: model selection, benchmarking against your real data, and honest answers about what the current state of the art can and cannot do for your problem.
Prototyping / Benchmarking / Model Selection / Due Diligence
Vibe-coded messes untangled, no judgement. When AI-generated code has grown past what anyone can safely change, the job is to work out what it actually does, pin that down with tests, and restructure it so people - and their agents - can build on it again.
Code Archaeology / Test Coverage / Refactoring / Documentation
Software for hardware: embedded firmware, motion control, sensor fusion, and the device-to-cloud plumbing in between. The work reaches down to the timing-and-interrupts level and up to the fleet dashboard, including the models that run on-device.
Embedded C / RTOS / Sensor Fusion / Edge Inference
Engineering of the calibre large institutions buy, sized for companies that are nothing like that big.
The work happens on your systems, with your data, under your constraints, next to the people who own them. What decides whether a project succeeds is rarely written in the brief, and you only find it from the inside.
The first milestone is the narrowest thing that tests the riskiest assumption against your real data. When an idea does not hold up, you find out in weeks and for a small fraction of the budget.
Proven infrastructure does the heavy lifting, and frontier models go where they change the outcome rather than where they decorate a slide deck. A good share of the best work here has been talking a client out of the expensive option.
Documented, tested, and deployed on infrastructure you control, with your team able to run and extend it after the engagement ends. No black boxes, and no dependency on anyone else's roadmap.
Procyon Research is an engineering consultancy with more than twenty years of industry experience building intelligent systems, embedded software, and the data infrastructure underneath them. The work spans industries because the discipline is the same everywhere: understand the domain and its constraints, measure the results, and ship production-grade systems designed for the people who will run them.
The large consultancies aim at institutions with thousands of staff and budgets to match. Procyon does work of that calibre for companies an order of magnitude smaller, without the overhead that usually arrives with it. Ambiguous data, entrenched legacy constraints, and workflows nobody has automated are the projects it takes on most gladly.