Kevin Matos · AI Work

Kevin Matos

Fractional CTO / AI. I use AI as the value-creation lever inside real businesses — and I ship, not slideware.

How I approach AI

Most AI work stalls at the demo. Mine ships into production and gets used. I pick problems where AI creates measurable business value, build the full system end to end, and hold it to product rigor — real evaluation, honest tradeoffs, and the unglamorous last mile that separates a working tool from a slide.

Selected work · client engagements anonymized

01 PRODUCTION · LEGAL AI

A private AI research system that never leaves the building

Problem
A solo law practice needed AI leverage on privileged case material — but privileged data can't touch the cloud, which put every mainstream AI tool off-limits for exactly the highest-value work. In legal, a fabricated citation isn't a UX annoyance; it's a malpractice risk.
Approach
I built a fully air-gapped system that runs on the firm's own hardware: a local open-weight LLM paired with hybrid retrieval — vector search plus a knowledge graph — so every answer is grounded in the actual case record. I selected the model through a formal bake-off scored on citation faithfulness (94.9%) and keyword recall (92.6%), not on vibes.
Outcome
Live on-prem, with an evidence-sweep capability and an operating guide the attorney uses directly in casework. No data ever leaves the office.

Signal: data sovereignty as product strategy · evaluation discipline · full-stack local deployment

02 PRODUCTION · WORKFLOW

Turning slow, leaky intake into instant qualified routing

Problem
A multi-state firm was losing leads to slow manual intake. Speed-to-response is the single biggest lever on lead conversion, and theirs was measured in hours.
Approach
I built an automated intake product that qualifies inquiries, routes each one to the right partner by jurisdiction, and fires instant notifications — white-labeled to the firm's own brand.
Outcome
A repeatable lead-generation engine rather than a one-off build — designed to be reused across future clients.

Signal: business-outcome framing · productization · speed-to-lead as the metric that matters

03 0 → 1 · SAAS

A job-application autopilot, built end to end

Problem
Applying for roles is high-friction and repetitive — easy to do, hard to do well at scale.
Approach
I designed and built the full product solo: a web app plus a browser extension that autofills applications with fuzzy field detection, backed by an LLM. Including the hard parts — the extension, and mapping messy real-world forms.
Outcome
Near-launch, owned end to end — product, backend, and the browser layer most people avoid.

Signal: zero-to-one ownership · full-stack execution · the unglamorous last mile

Also in flight

PaySwitch AI — a provider-neutral advisor for switching payment processors, drawn from payments domain depth built at Mastercard. A market gap I spotted, scoped to a phased MVP.

Capabilities

Product

$20M+ incremental revenue · Mastercard 0 → 1 product ownership evaluation & benchmark design exec-level tech assessment GTM strategy Account Level Management

AI & Systems

local + cloud LLM deployment RAG + hybrid retrieval vector databases knowledge graphs embedding-model selection model evaluation Chrome extensions Next.js / FastAPI / Postgres

Domain

payments & fintech legal tech SMB operations

Credentials

Generative AI with LLMs · DeepLearning.AI AI Product Management · Duke Claude Code in Action · Anthropic AI Fluency · Anthropic AWS ML Engineer Associate · in progress