What we do
Agents, vision and machine learning, built into real products.
Much of 5Swim's recent work is AI work. We build systems where models do real jobs — reading photos, answering domain questions, analysing imagery, forecasting — and we spend as much effort on the surrounding machinery: permissions, approvals, audit trails and the evidence that an answer is right.
That includes infrastructure for AI agents themselves: a gateway that governs what agents may do, a network where autonomous agents trade opinions with each other, and a specification language for checking code that AI has written.
An MCP gateway that sits between AI agents and a company's business tools. Every call is checked against YAML policies, risk-scored, held for human approval when needed, and recorded in a hash-chained, signed audit log — with a management console for the humans in the loop.
A machine-to-machine opinion relay for autonomous agents, written in Rust. Agents publish structured signals over a binary TLS protocol, the relay computes a consensus lean and settles outcomes — with no on-wire execution and a covenant every agent accepts on admission.
A golf tournament platform (mobile app, cloud functions and a provider portal) with an AI rules assistant: players photograph a situation on the course and get rules guidance grounded in the official rules through retrieval-augmented generation on Gemini.
A hybrid geospatial AI system that turns a golf-course name or location into a GIS and robotics dataset: it fuses OpenStreetMap priors with satellite imagery analysis, enforces topology rules and exports GeoJSON, KML, DXF and navigation grids.
View on GitHubPersonalised children's picture books: a short questionnaire becomes a fixed-length story illustrated from a single character description, previewed in the browser and sent to print only after payment is confirmed. Illustration, payment and print providers are swappable.
A nutrition tracker that analyses food photos with a vision model and logs the result to a daily intake, with trends over time.
An asset-centric research platform for algorithmic trading: engineered features, sixteen model types from gradient boosting to transformers and regime filters, backtests with Monte Carlo runs, and paper-trading hooks. A research and education codebase, not a live-trading product.
Smart contracts, wallets and ledgers with a job to do.
Putting the same tools to work on environmental problems.
Ideas we build to find out whether they hold.
Leave a message and tell us about it.