Jesús Martínez
ES

Work

What I’ve built for clients.

Each case study tells the whole story: the situation, the decisions I made along the way, and the numbers at the end.

The rest of the record

Everything else.

Shorter entries for the rest of my client work. Each one shipped and ran in production.

AI Engineering2023 to now

  • 2026
    A voice companion that answers in the creator’s own voiceSubscribers text an AI persona and get replies back as synthesized audio, built into an existing subscription platform.
    Audio subscription platformVoice AI
  • 2026
    A knowledge base built to be auditedMulti-tenant retrieval where every passage traces back to its source, built so another institution can inherit it.
    Training productRetrieval
  • 2026
    The product API behind a counter-speech trainerWrites evidence-backed counter-arguments and grades users’ own attempts, on top of the audited knowledge base above.
    Training productAgents
  • 2025–26
    An AI tutor that reacts to the roomThe orchestrator behind live tutoring sessions, responding to voice, cameras and motion sensors. Part of a platform team.
    Ed-tech platformMultimodal
  • 2025
    Policy answers without a support ticketA company-wide assistant that answers internal policy questions from the firm’s own documents.
    Legal services firmRetrieval
  • 2024–26
    One scheduling agent across four messaging appsCalendar management over Signal, Slack, WhatsApp and Telegram. A rebuild cut replies from 3 minutes to under 20 seconds and running costs by 90%.
    Productivity startupAgents
  • 2024–26
    300 emails a day, triaged and turned into tasksA classification and summarization pipeline for 50+ users, with every result stored as data you can query.
    Productivity startupDocument AI
  • 2024–26
    The agent backend of an inbox-and-calendar appOrchestration behind Tomo, an iOS assistant that turns email and schedule into a short list of priorities.
    Productivity startupAgents
  • 2024–25
    A bilingual assistant for five restaurant kitchensStaff ask about recipes, procedures and repairs in English or Spanish. Manager interruptions fell by about 75%.
    Restaurant groupAssistants
  • 2024
    Safari proposals from plain-English questionsA natural-language-to-SQL agent with schema checks and query validation. Trip planning time fell by 20%.
    Luxury travel agencyData agents
  • 2023–24
    Open models that match GPT-4 on contract riskFine-tuned Llama 3 and Mistral to run in isolated environments, so client documents never leave them. Part of an engineering team.
    Legal-tech platformFine-tuning

ML Engineering2019 to 2023

  • 2020–23
    Legal entity extraction, four times fasterReplaced slow production NER models with a new architecture that was also more accurate.
    Legal-tech companyNLP
  • 2019–22
    Pharma documents tagged better than human reviewersAn NLP tagger that beat reviewer accuracy by 10%, plus quarterly sales forecasts and cleanup for poor scans.
    Pharmaceutical companyNLP
  • 2019
    People counting from ordinary ceiling camerasFoot-traffic counts from overhead RGB cameras instead of dedicated sensors, running in several stores.
    Retail storesVision
  • 2019
    Height checks from depth sensorsAccess control that estimates a person’s height from RGB-D data, deployed at more than 10 venues.
    Entertainment venuesVision

Have a problem like these?