Gerson Azevedo — AI Engineer

About Me

Software engineer with 20+ years of experience and a two-time founder, now working in AI engineering. I build production systems side by side with non-technical stakeholders — psychologists, school administrators, government program managers — translating messy real-world requirements into shipped products. Currently the sole engineer at Nami, where I design, deploy, and evaluate a clinical AI agent used in live therapeutic settings. My background combines engineering with business (B.A. in Marketing), and I write about applied AI on my blog.

Read my full story →

Skills

Founder Experience: Talking to customers, creating presentations, pitching ideas for investment

Technologies: Python, Ruby on Rails, OpenAI Agents SDK, agentic/multi-agent systems, RAG/Agentic RAG, evals (LLM-as-judge + code-based), Anthropic/OpenAI/Etc APIs, LiteLLM, PostgreSQL + pgvector, Redis, Sidekiq, React/Next.js, TypeScript, Node.js, AWS, Docker

Highlights

Architected and built the full AI stack for a clinical mental health agent at Nami — multi-model routing, RAG pipeline, evaluation framework (evals), and WhatsApp integration in production with real patients.
Led the technical and operational execution for the Brazilian stage of the IOAI (International Olympiad in Artificial Intelligence) — 700,000+ registrations and 80,000+ exam submissions in Brazil.
Responsible for educational software used by 5,000+ teachers and 20,000+ students/parents across 70+ public schools in southern Brazil.

Experience

2025 – present

Nami

AI Engineer (sole engineer) — Clinical AI agent on WhatsApp, supervised by licensed therapists

Work directly with clinical co-founders (a psychiatrist and a psychologist) to translate therapeutic protocols into system behavior — running open and axial coding sessions on production traces with domain experts to define evaluation criteria and safety rules.
Designed and operate an agentic compound AI system orchestrating multiple models — proprietary (Anthropic, OpenAI) and open source — where a router evaluates each conversation and directs it to specialized models based on clinical risk level, with tool calling to escalate high-risk cases to human professionals.
Built an evaluation framework from scratch: 5 LLM-as-a-judge evaluators and 3 code-based evaluators, with synthetic conversation generation (~100 traces) and continuous production trace analysis.
Shipped the full AI stack end to end: multi-model routing (crisis detection vs. general conversation), RAG pipeline, prompt validation, and business rules.

Stack: Python, Anthropic API, LiteLLM, LLM Evals, LangChain, ChromaDB, HuggingFace, Phoenix (Arize), Ruby on Rails, Heroku

2019 – 2025

EduSpace

Co-founder & Lead Engineer — Edtech startup for public school networks

Deployed and operated a full school management platform (web, iOS, Android) across 70+ public schools — working directly with school administrators, teachers, and government program requirements (Bolsa Família) to serve 5,000+ teachers and 20,000+ students/parents across 5 access roles.
Led a team of 5 developers across web and mobile, handling infrastructure challenges including 100–200 req/s traffic spikes and the complexity of heterogeneous public school networks.
Designed and shipped a multi-platform olympiad system (web, iOS, Android) supporting online exams, printed answer sheet scanning, real-time ranking and analytics — used in the Brazilian stage of the IOAI (700,000+ registrations, 80,000+ exam submissions) and OBICT (500,000+ registrations, 50,000+ submissions).
Built an AI-powered messaging system that converts user queries into personalized SMS/WhatsApp messages using marketing triggers, now used for school olympiad communications.

Stack: Ruby on Rails, React, Next.js, TypeScript, PostgreSQL, Redis, Sidekiq, Heroku, AWS, GitHub CI, Google Maps API, OAuth

2018 – present

Servicient Co. — MyTennisLessons.com / FitnessTrainer.com

Contractor — One of the largest platforms for finding tennis lessons in the US (~50–100K monthly visits)

Integrated payment solutions (Plaid for trainer payouts, PayPal/Braintree for client payments), geocoding for training locations, and Elasticsearch for fast search across both platforms.
Led a full redesign of FitnessTrainer.com with a UX team, including landing page performance optimization.

Stack: Ruby on Rails, React, ActiveAdmin, PostgreSQL, Redis, Sidekiq, Heroku, Imgix, Filestack, Rollbar, New Relic

2018 – 2019

Pulsus

Early team member — Latin America's leading Mobile Device Management (MDM) platform (1M+ managed devices)

Served as both UI/UX designer and frontend developer in the early team, working directly with stakeholders to define and ship features.
Built a React interface for truck driver route tracking and comparison against optimal routes for fuel cost analysis.

Stack: React, Ruby on Rails

2015 – 2018

Porta-Jaleco

Co-founder — D2C e-commerce for healthcare accessories, with its own product line and outsourced manufacturing

Took a physical product from concept to market: designed the product with input from students across 11 universities, managed outsourced manufacturing and quality control, and ran digital marketing — reaching ~50 orders/day (peaks of 80–100) with a 4.8/5 customer rating. Selected by ESPM-Sul's business incubator (i-LAB).
Built and automated the entire operation: custom e-commerce with full order-cycle automation (payment processing, electronic invoicing, carrier tracking), plus a B2B ordering portal and CRM designed for a retail representative network.

1998 – 2014

Earlier career

Freelance web development, agency work (full-stack: HTML, CSS, JavaScript, PHP, SQL), C#/.NET contracting (cost estimation tools, ERP systems), teaching web development and advanced Excel/VBA at Senac-RS, and building a Rails-based remote voting platform for advertising awards (ESPM).