Kevin Street
I build practical AI and data systems for financial and professional-services teams — and the evaluations that show whether they are ready for production.
I work at the point where an AI idea has to prove itself: document extraction, grounded retrieval and tool-using workflows that cite their sources, keep a person in the loop, and ship with the evaluations that say whether they are ready for a pilot. The question I put to every system is the one a risk committee asks — how would we know if this is wrong? Backtesting is offline evaluation, monitoring is observability, limits are guardrails.
Projects
Working applications, each documented end to end: the problem, the architecture, a representative test set, repeatable evaluations and a record of what changed between versions. Numbers appear as soon as the eval suite runs; until then the page shows the plan, honestly labeled.
| Project | What it does | Evidence |
|---|---|---|
| EEOC charge intake: LLM extraction with source spans and human review | Turns an EEOC charge PDF into a reviewed, structured intake record — every field cites the page and text it came from, and nothing is exported until a person approves it. | In progress |
| Grounded Q&A over approved employment-law sources, with abstention | A retrieval assistant that answers only from an approved corpus, cites the passage inline, and says "not supported by the approved sources" instead of guessing. | In progress |
| New-matter intake agent with scoped tools and approval gates | An agent reads an intake email, runs a conflicts search with read-only tools and drafts the matter — but cannot open it until an approver signs off, enforced server-side. | In progress |
What I do for teams
| Prototype to pilot | A working application in weeks, not quarters: assistants, document extraction, knowledge retrieval and workflow tools on your approved models and data, with the interface, logic and data connections built as one small, testable solution. |
|---|---|
| Evaluation you can defend | Representative test sets and repeatable evaluations for accuracy, unsupported outputs, failure handling, response time and cost — the evidence a steering committee needs before approving a pilot. |
| Data pipelines that hold up | Python and SQL pipelines, monitoring and reporting built for regulated environments: audited inputs, reproducible runs, and alerts that fire before anyone has to notice. |
Every engagement ends with documentation a successor could run from: setup, dependencies, data flows, test results, and the work left to make it a supported production service.
Work with me
Building AI or data systems and need someone who can ship the prototype and prove whether it works? Email info@kevindstreet.com with a sentence or two about the problem. I reply within two business days. Also on LinkedIn and GitHub.