Open to work

Diego Costa AI engineer for industries that still run on spreadsheets.

Construction bidding at work, fitness tooling on my own time, six years of ad operations before either — I put AI inside products for people who never asked for AI, and the win isn’t the model. It’s the hours it hands back.

View work Get in touch
Montevideo, Uruguay UTC-3

About

I got here by automating my own job.

Diego Costa working on his laptop

For six years I worked in ad operations. Plenty of the job was repetitive by design, so I started automating my own part of it — scripts first, then small tools — until the work got done faster and my time went to the decisions instead of the clicking. Automating turned out to be the part I wanted to do full time, so I left, finished the degree I had started while working, took the freelance jobs I could get, and came out the other side a developer.

Since mid-2025 I’ve been building products with an LLM inside, all day. The model is rarely the hard part. The hard part is everything around it: grounding an assistant in the manual so a construction estimator can check the answer instead of trusting it, making sure a client’s documents stay inside their own tenant, and being able to tell what the thing actually did at 3am when it misbehaved. That’s the work I do on AnswerAgent, on SBI Bids, and on my own MCP tooling.

I also write the manual. On SBI Bids I documented the product screen by screen and validated it against the code across all eight workflows — because a feature nobody can explain isn’t shipped, it’s just deployed.

I build from Montevideo through my own company, Caivanode SAS, mostly alongside teams in the US, and my working language on the job is English: the pull requests, the docs and the code. What I’m looking for now is a product team to build with — one where AI is the means and the product is still the point.

3
Products I’ve shipped to production
23
MCP tools published on npm
6
Years in ad ops before software

Work

Some of what I’ve made

Two products in production at one company, one published tool, and the thesis where I first put AI in front of someone who wasn’t me. SBI Bids is the client’s repo and stays private — that one I describe, I don’t show. The rest you can open.

SBI Bids

Bid comparison and AI assistant platform for commercial construction

Estimators compare vendor bids that arrive as PDFs and spreadsheets, and the numbers end up in a Schedule of Values — the Excel workbook the whole bid rides on. I write the business rules inside it: insurance rates, GC/GR multipliers driven by project duration, net-outs between trades, accepted alternates and exclusions. I also wrote the user manual, and built the tool that answers with the manual in hand — a link to the exact section, and when the manual doesn’t cover it, it says so. Private client repo, hosted on Azure.

  • Schedule of Values rules
  • Answers link to the manual
  • Private client repo
  • TypeScript
  • React
  • PostgreSQL
  • Anthropic
  • Excel
  • Vitest
  • Azure

AnswerAgent

Production AI agent orchestration platform for enterprise clients

An agent orchestration platform built on Flowise, running on AWS Fargate. I work on the parts that decide whether an agent can be trusted in production: tenant isolation on the document store, runtime guardrails on the public embed endpoint, and the five-second timeout that turned out to be a no-op because the MCP subprocess never inherited its env. Tracing on every call, so when it misbehaves at 3am there is something to read.

  • TypeScript
  • Node.js
  • React
  • PostgreSQL
  • MCP
  • RAG
  • Langfuse
  • Fiddler
  • AWS
My 70 merged PRs Repo

hevy-mcp

Published Model Context Protocol server for the Hevy fitness API

A public, CI-backed MCP server exposing the Hevy API as 23 strongly-typed tools over stdio for Claude Desktop, Cursor and VS Code. Zod edge-validation with self-correcting errors, dry-run-by-default writes, and schemas auto-generated from Hevy’s OpenAPI spec via a weekly GitHub Actions workflow that opens PRs on drift.

  • TypeScript
  • MCP
  • Zod
  • GitHub Actions (OIDC)
  • Vitest
npm GitHub

SmartLift

University thesis — AI workout-generation API

My thesis project at Universidad ORT: a Ruby on Rails backend that generates personalized workout routines with AI, plus user tracking and analytics. The first thing I built where the AI had to hold up in front of someone other than me.

  • Ruby on Rails
  • PostgreSQL
  • AI
  • REST API
GitHub

Experience

Where I've been building

Six years in ad operations, then software. The order matters — it’s why I know what a spreadsheet costs the person who has to fill it.

  1. Jun 2025 — Present

    AnswerAgentAI

    AI Engineer

    Remote · US team · contracting through Caivanode SAS

    Engineer across three products in the same engagement: AnswerAgent, an AI agent orchestration platform (Flowise, AWS) for enterprise clients; SBI Bids, the construction-bid product, a separate codebase on Azure; and an internal desktop app that shipped and was later discontinued.

    • Write the business rules inside the Schedule-of-Values export on SBI Bids — the Excel workbook the whole bid rides on: insurance rates, GC/GR multipliers driven by project duration, inter-trade net-outs, accepted alternates and exclusions — and I run the acceptance pass the client signs off on, PII-scrubbed, gate by gate.
    • Wrote the user manual the assistant now cites — the eight workflows, screen by screen, checked against the code — and built query_manual, the tool that answers “how do I do X” straight from that manual with a link to the exact section, and says the manual doesn’t cover it instead of inventing UI.
    • Closed the multi-tenancy gap in the document store — chunks were being written with no owner id, so tenant filters couldn’t match them and users couldn’t see their own documents — and put Fiddler’s runtime guardrails on the public embed endpoint, which had been skipping them because embed requests carry no user.
    • A client’s agent was timing out at five seconds, intermittently. The timeout everyone had configured was a no-op — the MCP subprocess never inherited the env — and behind it the vectorstore query ran one unindexed lookup per result document, so under load the tail crossed the limit. Fixed both.
    • Got the company’s Electron desktop app shipping again on macOS and Windows: signed builds had been failing notarization for a month, and auto-update published a single manifest for two architectures, so arm64 users updated into the wrong binary. Then built the production monitoring that had never existed behind it — the CPU, memory and disk alarms had sat in INSUFFICIENT_DATA since March because nothing was installed to emit the metrics.
    • TypeScript
    • RAG
    • MCP
    • Langfuse
    • Fiddler
    • AWS
    • Electron
    • Docker
    • PostgreSQL

  2. 2024

    Freelance

    Full Stack Developer

    Montevideo · Remote

    A short engagement building the Deal Management module for a DSP ad platform — the same kind of platform I had spent six years operating, now from the other side of it.

    • Angular front end over a Node API on MySQL: the screens where an ad-ops team sets up deals and the endpoints behind them.
    • Knowing the job from the inside was the whole advantage — I had been the person filing the tickets against a module like this one.
    • Angular
    • Node.js
    • MySQL

  3. Jul 2017 — Dec 2023

    Entravision

    Ad Operations · Programmatic

    Montevideo

    Six years running campaigns on a programmatic advertising platform — and automating my own part of the job until automating became the job.

    • Automated internal ad-ops processes with custom scripts — the work that turned me into a developer.
    • Built the backend tooling the DSP ran on (Node.js, SQL).
    • Built the interactive ad units that shipped in Spotify campaigns (React, HTML5, CSS3).
    • Was the person who explained the platform to the regional teams that had to use it — the habit that later turned into writing the SBI Bids manual.
    • React
    • Node.js
    • SQL
    • Ad Ops

Education

  • 2022 — 2025

    Universidad ORT Uruguay

    Programmer Analyst (Analista en Programación)

  • 2021

    CEIBAL · Jóvenes a Programar

    Web Development Diploma

Stack

Tools I build with

Grouped by what I built with them, not by how well I think I know them. If something shows up in three places, that’s your answer on how often I use it.

Contact

Looking for a product team.

Remote, on a team that ships — as a contractor through my own company, or as a full-time hire. Either way, write to me and ask me about anything above. It’s all checkable.

Email me CV (PDF)
diecoscai@gmail.com

Montevideo · UTC-3 — one to two hours ahead of New York depending on the season: the full East Coast day, and West Coast mornings through early afternoon