AI Engineer

Lara Mateo Portfolio

Tech Lead & AI Product Engineer; Founder, Kenda

A "case studies" section instead of a project list — each entry names the problem, the constraints, and the outcome, backed by a live AI FinOps startup and two open-source MCP servers with real GitHub stars and contributors.

Next.jsTypeScriptModel Context ProtocolVercel AI SDKClaude API

Location

Buenos Aires, Argentina

Companies

KendaClarity RCMReflexAIMercado LibreGlobalLogic

Skills

Next.jsTypeScriptModel Context ProtocolVercel AI SDKClaude API

Notable

  • Founded Kenda, an AI FinOps platform live at kenda.app with SOC 2 Type II in progress
  • Marksight (open-source Markdown-to-Claude-Skill editor): 18 GitHub stars, 11 contributors, 100+ users
  • Umbryn MCP: open-source PHI/PII redaction server published to PyPI
  • Invited closing speaker, Café Cursor #4 (Buenos Aires, August 2026) on building MCP servers
  • Went from Senior Engineer to Tech Lead in ~6 months at Clarity RCM, a HIPAA-regulated healthcare platform, leading a team of up to six

What makes it work

A breakdown of the choices that make this portfolio stand out.

1

Case studies framed as problem → constraints → what I did → outcome

Instead of a screenshot-and-tech-stack project grid, the two featured case studies (the Tech Lead promotion, the Core Web Vitals work at ReflexAI/Mercado Libre/GlobalLogic) are written as structured narratives with an explicit "what I took from it" reflection. That last section is unusual — most portfolios stop at the outcome and skip the self-assessment.

2

Side projects that form a visible thesis, not a grab-bag

Umbryn redacts PHI/PII before it reaches an LLM, Marksight turns documents into Claude Agent Skills, Kenda reconciles what AI agents actually cost — three different projects that all attack the same problem (the unglamorous infrastructure agentic AI needs before it is trustworthy in production). The GitHub README states this pattern explicitly rather than leaving it for the visitor to notice.

3

Verifiable open-source footprint, not just claimed skills

The Umbryn and pls-touch-grass MCP servers link to live PyPI/npm packages and GitHub repos with actual commit history; Marksight shows 18 stars and 11 contributors on GitHub. The GitHub profile independently corroborates the site's claims (same founder title at Kenda, same pinned repos, cross-links back to laramateo.com) rather than the portfolio being the only place the claims exist.

4

A talk that admits the demo broke

The "Building MCPs in Cursor" talk writeup and the linked blog post don't sand off the failure — the agent got frustrated mid-demo, started editing the presenter's code, and broke the MCP live on stage. Leading with that instead of a polished highlight reel is a distinctive, checkable (YouTube link, GitHub demo repo) piece of content most portfolios wouldn't include.

What AI Engineers can take from this

Specific, actionable tips to apply to your own portfolio — no generic advice.

  • Write project write-ups as problem → constraints → what you did → outcome, and add a closing "what I took from it" line — it reads as reflection, not a résumé bullet restated.

  • If your side projects share a theme, say so explicitly in one sentence. A reader forgives three portfolio projects; three projects that are visibly one thesis reads as intentional expertise.

  • Link every claimed tool or package to its live PyPI/npm/GitHub page. A named MCP server with no working link is a claim; with one, it is a fact anyone can check in ten seconds.

  • Cross-link your portfolio and GitHub profile to each other, and keep the bio consistent (same title, same current company) on both. That consistency is what makes a corroborating link actually corroborate.

  • A demo that goes wrong is more memorable and more credible than one that goes perfectly — if you have one, write it up instead of hiding it.

Projects

All of Lara Mateo's side projects tackle one gap: infrastructure AI agents need before a regulated team can trust them. The cards cover PHI redaction, cost reconciliation, agent-ready documents, and an MCP demo that broke on stage, written up in public.

Open SourceadvancedJuly 2026; v0.2.1 on PyPI

Umbryn MCP

PythonModel Context ProtocolMicrosoft PresidioRuffGitHub Actions

A self-hosted MCP server with three tools: redact, restore and detect. It swaps PII and PHI for reversible placeholders before text reaches an LLM, keeping the token map local. It recognizes HIPAA identifiers (NPI, DEA, Medicare MBI, MRN, CLIA) and fails closed, blocking the request when detection is uncertain.

It solves a real blocker for LLMs in healthcare, sending PHI to a model provider, with a design that fails safe rather than fails open.

  • Reversible placeholders; token map never leaves the machine

  • HIPAA identifiers such as NPI, DEA, MBI, MRN and CLIA

  • Fail-closed: uncertain detection blocks the request

Open SourceintermediateStarted September 2025; ongoing

MarkSight

Next.jsReactTypeScriptCodeMirrorMermaidTailwind

A browser-only Markdown workspace with live side-by-side preview, an outline, keyboard shortcuts and local-only storage. It exports to HTML or PDF, and with one click it packages any document as a SKILL.md bundle for Claude and other agents. It also exposes an MCP server for Claude Code.

A real open-source project with outside contributors and a feature (document to Agent Skill) that fits how teams now use AI tools.

  • One-click export of any document as a Claude Agent Skill

  • 13 contributors and 184 commits

  • Local-only storage, nothing leaves the browser

Creative DemointermediateAugust 2026; demoed at Café Cursor #4, Buenos Aires

Please Touch Grass MCP

TypeScriptNode.jsModel Context ProtocolOpen-MeteoOpenStreetMap Overpass

A joke MCP server with five tools: frustration_detector, where_to_touch_grass, grass_conditions, touched_grass and gotta_go. When you rage at your agent, it chains weather and park lookups to send you outside. It needs no API keys and stores state in a local JSON file.

It shows, and writes up honestly, that a tool description works as a prompt steering agent behavior. That's practical MCP knowledge taught to a live room.

  • Five tools; the frustration detector fires from its tool description alone

  • Keyless APIs: Open-Meteo weather and OpenStreetMap parks

  • Public write-up of the demo breaking on stage

SaaS ProductadvancedOngoing; invite-only

Kenda

Next.jsTypeScriptFastAPIPythonAWS LambdaDynamoDBAmazon Bedrock

An AI-spend reconciliation platform. It attributes LLM usage by sub-agent, agent, user, team, workflow or API key, then reconciles measured spend against the provider invoice at day, provider and model granularity. It uses read-only keys and doesn't store raw prompts or completions.

Founding a FinOps product for agent spend means working on both the engineering and the finance sides of AI adoption, with security constraints designed in from the start.

  • Spend sliced six ways, from sub-agent to API key

  • Measured-vs-billed reconciliation at day, provider and model level

  • Read-only keys; no raw prompts or completions stored

Ready to build your portfolio?

Follow the AI Engineer roadmap — skills, projects, and timeline to get hired.

AI Engineer Roadmap