gofri Why two g's? When I was a kid my grandma called me Gofri, and the family picked it up. She said the G hard, almost twice, so I write it twice.
Julian Ramirez · Senior Product Engineer · Applied AI

Vague customer problems, shipped to production.

I work with customers to find what’s worth building, make the product and architecture decisions, and ship applied AI systems.

Email me CV ↓ Selected work ↓
customervague, real, urgent
teachers aren't using itTicmasone prompt, getting worseMercado Libre
decidewhat to build and what to skip
interviews: teach by curriculum topicTicmasevals show it regressing as it growsMercado Libre
buildcontracts, evals, pipelines
recommendations organised by topicTicmassplit: retrieve, reclassify, answerMercado Libre
measuredid it change behaviour?
50 → 1,000+ teachers activated, cumulativeTicmas≈25% of steps bot-handledMercado Libre

WORK — 01

Selected work

production

Curriculum-aware recommendations

Ticmas · 2023–24

  1. askTeachers were not using the recommendations.
  2. foundMy first bet, personalising from usage and engagement data, barely moved anything.
  3. builtWith no dedicated ML engineer, led the technical experiment: pipeline, notebooks, and models on Bedrock.
  4. shippedCumulative activated teachers rose from about 50 to over 1,000 by the month after rollout, and the redesign reached every Ticmas customer.
Open case file →
v1 · one generic list
  • Recommendations should follow the topic I am teaching.
  • I want resources for the level of my class.
  • I do not want to search the whole catalogue.
  • Fraction stripsFractions
  • Story mapsReading
  • Food chainsEcosystems
  • Reading dataFractions · Ecosystems
about 501,000+ startmonth after rollout

Pilot: about 50% lower 30-day inactivity than non-pilot schools (observational).

From a generic list to curriculum-aware recommendationsreconstruction, illustrative cumulative
in ongoing use

Manchester City fan-verification dashboard

Manchester City · via Veritran · 2026

  1. askA dashboard over fan-verification data, with no defined scope.
  2. foundThe records already existed. What was missing was a view of what happened to them: verified, pending, failed or duplicated.
  3. builtTook a request with no defined scope and owned the whole implementation, with the product and technical decisions.
  4. shippedFirst version in two weeks, final in one month, still in use. My point of contact reported that its country view informed enabling onboarding in two more countries.
Open case file →
60records, no view verified 36 60.0% pending 14 23.3% failed 6 10.0% duplicated 4 6.7%
  1. Snowflake prototype · validated by stakeholders
  2. v1 · 2 weeks
  3. final · 1 month
  4. ongoing use
60 records, four outcomessynthetic data pending ≠ abandoned
production

Support-bot orchestration

Mercado Libre · 2021–22

  1. askA support-bot prompt that was getting worse as it grew.
  2. foundEvaluations showed the prompt regressing as it grew.
  3. builtInformal technical lead alongside two senior developers.
  4. shippedAbout 25% of support handling steps were bot-handled, as reported by the manager and in performance reviews.
Open case file →
  1. You are a support assistant.
  2. Answer in the customer's language.
  3. If the order is late, explain the carrier states.
  4. If the customer asks for a refund, apply the policy.
  5. If the message is unclear, ask one question.
  6. If unsure, hand over to a person.

evals flag regressions as the prompt grows

  1. retrieval
  2. reclassification
  3. answering
human escalation
One prompt, split into stagesillustrative prompt lines

More work

More work
periodwherearrivedshippedstage
2026 Veritran A request for a bank-data product that clients could evaluate. Delivered in one month for technical evaluation and demos. Demo stage, no production customer. demo stage
2026 Veritran A recommendation app that took 10–20 seconds to paint. Lighthouse cold-load FCP under 200 milliseconds. Self-reported; I owned this change only. demo stage
2024–25 Anonymous agency via ggofri An alternate-reality game that needed agents behind it. Four systems, built personally: journalist/newsletter, clue generation, investigator workflow and simulated players. Players were simulated. production
2024–25 Live Nation via Globant Several external providers, each with its own shape. One normalization layer in front of more than five providers, with retries, mismatch alerts and CI tests. production
2024 Ticmas Two teams and a stakeholder gap. Managed 15 direct reports (1 QA, 2 designers, 12 developers), Apr–Nov 2024. Developed and sponsored two official Staff promotions. production
2024 Ticmas A backend that had to move without stopping. More than 80% of the backend migrated by Nov 2024, including all important endpoints. production
2023–24 Ticmas Students needed help understanding, not direct answers. Designed the student-support workflow and implemented the Bedrock integration. limited production
2023 Batch via eagerWorks A chat that customers felt was broken. Diagnosed in one week, fixed over a two-week sprint, rolled out in stages. production
2022–23 Expedia (anonymous partner) via Globant Existing component endpoints, no combined flow. Combined travel-package flow in production for an anonymous partner. production
2019 Independent Small shops that sold over WhatsApp. Two paying entrepreneurs used it for approximately one year. No sales or revenue totals tracked. retired

TIMELINE — 02

Timeline

Jobs

  1. Lenovo Lenovo · global marketing team, via a staffing company
    Software Engineer
    May 2021 – Nov 2021
  2. Mercado Libre Mercado Libre
    Software Engineer
    Nov 2021 – Jul 2022
  3. Expedia Expedia · via Globant
    Senior Software Engineer
    Jul 2022 – Mar 2023
  4. eagerWorks eagerWorks · Batch
    Senior Software Engineer
    Mar 2023 – Sept 2023
  5. Ticmas Ticmas · earlier: Lead Engineer → Tech Lead
    Engineering Manager
    Oct 2023 – Nov 2024
  6. Live Nation Live Nation · via Globant
    Senior Software Engineer
    Dec 2024 – Dec 2025
  7. Veritran Veritran
    AI Product Engineer
    Jan 2026 – now

Practice

  1. ggofri ggofri
    Founder & AI Engineer
    Sept 2021 – Jan 2026

Side

  1. Independent projects Independent projects
    Side work
    Mar 2019 – Apr 2021
  2. Bosun Bosun
    Technical Founder (part-time)
    Feb 2026 – now
  1. Mar 2019 – Apr 2021 Independent projects Side work · Side
  2. May 2021 – Nov 2021 Lenovo · global marketing team, via a staffing company Software Engineer · Jobs
  3. Sept 2021 – Jan 2026 ggofri Founder & AI Engineer · Practice
  4. Nov 2021 – Jul 2022 Mercado Libre Software Engineer · Jobs
  5. Jul 2022 – Mar 2023 Expedia · via Globant Senior Software Engineer · Jobs
  6. Mar 2023 – Sept 2023 eagerWorks · Batch Senior Software Engineer · Jobs
  7. Oct 2023 – Nov 2024 Ticmas · earlier: Lead Engineer → Tech Lead Engineering Manager · Jobs
  8. Dec 2024 – Dec 2025 Live Nation · via Globant Senior Software Engineer · Jobs
  9. Jan 2026 – now Veritran AI Product Engineer · Jobs
  10. Feb 2026 – now Bosun Technical Founder (part-time) · Side

Periods overlap where side work ran alongside a job. They are shown as recorded, never added up.

ggofri · 2021–2026 · alongside full-time roles

ggofri, the practice.

>8AI systems across an ARG agency and four startups

A growth-engineering practice spanning AI delivery, architecture and startup team formation. I owned client strategy and hired freelancers for startup delivery; the agency's ARG systems I built myself.

BOSUN — 03 · part-time, alongside my full-time role

Parallel agents need one shared board.

todo

task-142agent-a
task-143agent-b
task-144

in_progress

done

Two agents, two different tasks. No collision.

Logbook keeps task and kanban state on the filesystem, so several coding agents can ask for their current task and each get a different one. It is one of five tools I published at Bosun, all built on Ohtools, my TypeScript framework where agents explore a tool before running it. Each tool exposes the same definitions as a CLI and an MCP server.

Logbook project · Ohtools · task names illustrative

5tools published on npm · sole contributor

Research and selected projects

  1. preprint

    The Intervention Penalty: A Simulation Study of Human Checkpoint Costs in AI Coding Governance

    Julian Ramirez · Sofia Plajutin · April 2026

  2. preprint

    Artificial Intelligence in Cybersecurity: Reproducible Evidence on Vulnerability Discovery and Security Operations

    Sofia Plajutin · Julian Ramirez · April 2026

  • ohtools

    One inspectable tool registry for CLI, MCP and documentation.

    repo ↗ · v0.2.0 · last commit 2026-07-01

  • logbook

    Filesystem-based task and kanban state, exposed through MCP and a CLI.

    repo ↗ · v2.1.0 · last commit 2026-06-28

  • fake-end

    Independent project: local API mocks from YAML and TypeScript, including failure states.

    repo ↗ · v2.2.1 · last commit 2025-08-20

  • plot

    Layered specifications with inheritance and exact-target approval gates.

    repo ↗ · v0.2.0 · last commit 2026-06-28

What I stopped building

Crew: four months of work.
Then a decision to stop.

I spent about four months building Crew, then benchmarked it myself. In that comparison, the extra cost came without an automated correctness advantage. I shut the project down.

Read the project note →
Recorded tokens
+22%
Median machine time
+54%

No automated correctness advantage.

Small synthetic comparison · self-reported.
Not a production evaluation.

Latest writing

AI security after discovery: capability is not deployment confidence

An October 2026 evidence update on AI security: why vulnerability discovery, shipped fixes and control failures require different deployment judgments.

 All writing →

Looking back

I start from the customer's vague problem and stay until it runs in production. Each case below says what was mine, what was not, and what I could not prove. Open to senior product engineering roles: customer-facing, applied AI, with forward deployed work as part of the job.

3 cases, each one a piece. Open a case or try its demo to collect it.

  1. Curriculum-aware recommendations Cumulative activated teachers rose from about 50 to over 1,000 by the month after rollout, and the redesign reached every Ticmas customer.
  2. Manchester City fan-verification dashboard First version in two weeks, final in one month, still in use. My point of contact reported that its country view informed enabling onboarding in two more countries.
  3. Support-bot orchestration About 25% of support handling steps were bot-handled, as reported by the manager and in performance reviews.
  4. your problem
    Email me Download CV