Skip to content
Gairo Peralta
English Español
gairo@berkeley.edu

About

An engineer for problems that live on a map or in a document.

I'm Gairo Peralta, an AI/ML engineer. I build systems that turn city blocks, legal files and medical questions into answers people can check, then measure whether they hold up.

Focus
LLM agents, GeoAI, production ML
Education
BA Data Science, UC Berkeley · Master's in Applied AI, Tec
Currently
Lead AI Engineer, BlackPrint Technologies
Languages
English and Spanish
Portrait of Gairo Peralta

Gairo Peralta

Lead AI Engineer

The story

As Lead AI Engineer at BlackPrint Technologies, I design territory-intelligence products on Mexican census and commercial data. I led mireporte.ai, which turns a point on a map into a 25-section market study across eleven metro areas, and built field-sales routing, payment-acceptance scoring for ~5.5M merchants and a pipeline that reconciles 230k+ points of interest.

From 2025 to July 2026 I also led AI work for the Law Offices of Manuel Solis as a consultant, guiding a team of interns. Our LLM agents cut turnaround on T-Visa declarations and cover letters from about two hours to about fifteen minutes per document. We also built a meeting-intelligence platform with role-based access and a certified-translation pipeline on Gemini Vision.

From 2025 to January 2026 I was a visiting student researcher in medical AI at UT San Antonio (volunteer), where four agents reasoning over PrimeKG reached 54.69% accuracy on medical QA, a 24.24-point gain. I hold a BA in Data Science from UC Berkeley (2022) and am completing a Master's in Applied AI at Tecnológico de Monterrey (expected 2027, grade 98/100).

I work in English and Spanish, which helps when the data comes from INEGI and the client drafts U.S. immigration filings. I'm known professionally as Gairo Peralta; my full legal name is Gairo Yostin Peralta Alvarez in Mexico and Gairo Yostin Peralta in the United States.

Timeline

The path so far

Roles, research and degrees since 2022, newest first.

  1. 2025 – Present

    Lead AI Engineer

    BlackPrint Technologies

    Architected LLM-powered GeoAI platforms for 2.5M+ urban blocks, multi-agent systems (LangGraph), real-time AI sales agents (Next.js 15 & WhatsApp Cloud API), production monorepos, and high-performance geospatial matching pipelines.

  2. 2025 – Jul 2026

    Lead AI/ML Engineer (Consulting)

    Law Offices of Manuel Solis

    Developed custom LLM agents for legal documents and a meeting intelligence platform with RBAC.

  3. 2025 – Jan 2026

    Visiting Student Researcher

    UT San Antonio

    Medical AI research with knowledge graphs and multi-agent evaluation pipelines.

  4. 2024 – 2025

    AI Engineer

    PRODENSA (Contract)

    Built OCR and Llama 3.3 extraction pipelines with FastAPI + React/Vite for real-time PDF processing.

  5. 2023 – 2025

    AI/ML Engineer

    STRTGY

    Built an enterprise RAG assistant (Claude 3.5 Sonnet + FAISS), GPT-4 Vision competitor mapping and adaptive demand forecasting.

  6. 2022

    Graduated from UC Berkeley

    Completed a Bachelor's degree in Data Science, focusing on machine learning and statistical analysis.

  7. 2022 – 2023

    Data Scientist

    BlackPrint Technologies

    Delivered geocoding analytics and pandas/SQL ETL pipelines for property intelligence.

Principles

Working principles

01

Evaluate before you demo

A demo proves one answer; an evaluation proves the rate. The medical QA work reports 54.69% accuracy and a 24.24-point lift, not a highlight reel.

02

Check the cost before the query

On a ~79 GB BigQuery lake, a vague question can get expensive. In Meridian, a SQL validator and four-tier cost gates check every generated query before it can run.

03

Privacy goes in the pipeline

NetAiCore analyzes chatbot conversations under Mexico's LFPDPPP: anonymization runs inside the pipeline, and the shareable dashboard shows no PII.

04

Leave room for a human

BrokerAI's agent qualifies leads, but operators watch conversations live and can take over, with Telegram alerts when a chat needs a person.

Method

How a project runs

  1. 1

    Frame the decision

    Start from the decision the system should support and the data that already exists, not from a model.

  2. 2

    Test on your data

    A small prototype on your data, scored against an evaluation set, shows what works before anything scales.

  3. 3

    Build for production

    APIs, infrastructure as code, cost controls and access rules go in early, so the prototype can grow instead of being rewritten.

  4. 4

    Hand off and measure

    Documentation and training for the team that inherits it, plus the agreed metric, tracked after launch.

Exploring

What I'm exploring

  • 01

    Text-to-SQL that knows its limits

    Agents that check their own SQL, explain their joins and say so when the data can't answer the question.

  • 02

    Places as embeddings

    Representing blocks, POIs and neighborhoods as vectors, so models can compare places the way they compare text.

  • 03

    Graphs that ground LLM answers

    Multi-hop reasoning over knowledge graphs like PrimeKG, so every answer can point back to its evidence.

  • 04

    Analytics that protect people

    Getting cohorts, topics and sentiment out of conversations without exposing the people who wrote them.

Bring the problem. I'll bring the questions.

Send a short note about the problem and the data behind it. I reply within 24–48 hours.

Get in touch See the work