About
I build practical AI and data systems for the people who do skilled work — professional-services teams, small businesses and the trades — and the evaluations that show whether they are ready for production.
I'm Kevin Street. I build AI and data systems and document them like evidence: the problem, the constraints, the architecture, the test set, the numbers, and what changed between versions. This site is the portfolio. Everything on it is my own work, built and evaluated by me, so a hiring manager or client can judge it from the pages rather than from my say-so.
How I work
- Start with the question an owner or a risk committee asks: how would we know if this is wrong? If a prototype can't answer it, it isn't done.
- Early tests run on synthetic or approved sample data. No real records until there is a reason and a control.
- Anything that acts gets the narrowest tools that can do the job, a human review point before the consequential step, and an audit log of every call.
- Every system ships with a representative test set and repeatable evaluations for accuracy, unsupported outputs, failure handling, response time and cost — and a changelog that records what each version changed in those numbers.
- Documentation a successor could run from: setup, dependencies, data flows, test results, and the work left to make it a supported production service.
Why evaluation comes first
Risk management and machine learning are the same discipline with different names. Backtesting is offline evaluation. Monitoring is observability. Limits are guardrails. I hold every prototype to the standard a risk committee would hold a model to: it earns trust with evidence, not with a demo.
Let's Learn Out Loud
Separately, I teach what I build. Let's Learn Out Loud explains practical AI and data systems through real, open-source workflows: the chunking, embeddings, vector stores, LLM APIs and pipelines behind systems like the ones here, worked through step by step for people building their own.
Stack
Python · SQL · FastAPI · TypeScript · scikit-learn · Snowflake · AWS · Azure · Docker · Airflow · Git · Tableau · RAG and vector databases · LLM APIs (Claude, GPT, Gemini) · model validation · quantitative methods