Intercorp · structure
45.6% → 100%
conformance in production before the fix; 8/8 test fixtures after it
AI & Backend · Milan
Open to AI Engineer roles · EU
Open to AI Engineer roles · EU
I ship LLM systems
that survive production.
Eight years in backend; the last four shipping multi-agent, RAG and eval pipelines to real users.
Senior Software Engineer, AI & Backend. Eight years building backend systems — the last four shipping multi-agent, RAG and eval pipelines to real users. Measured, grounded, cost-aware. Based in Milan.
Now
2025 – 2026
Reeve · merge campaign
110+ PRs · 30 h
queued PRs cleared across 4 repos with parallel Claude Code agents
MindFortress · cost
~$16/day
unattributed AI spend traced to a stale staging worker and guarded in code
Wordbox · 10+ markets
500K+ users
served by LLM transcription and translation pipelines
- 45.6→100%production before, 8/8 fixtures after
- 110+ PRs · 30hqueue cleared with parallel agents
- ~$16/dayunattributed AI spend, traced and guarded
Selected work
Four problems, and what changed
How I approached each one, in the order it happened. Every case reads in about ninety seconds.
01 · Intercorp · AI Engineer
Evals before vibes: fixing a prompt that drifted
Before: 45.6%. After: 100%.
LLM-as-judge fixtures across 5 dimensions and 19 items turned prompt changes into a measured release gate. Structural conformance: 45.6% in production before the fix, 8 of 8 test fixtures after it.
- Gemini 2.5 Pro
- Spring AI
- LLM-as-judge
02 · MindFortress · Tech Lead
Finding the AI spend nobody could explain
Before: ~$16/day. After: guarded.
A spend bucket with no owner was costing about $16 a day. I traced it to a stale staging worker running against the production database, shipped a guard in the code and split the production and staging credentials.
- Python
- Railway
- Root-cause analysis
03 · Reeve · Tech Lead
110+ queued PRs cleared in 30 hours
Before: 110+ queued. After: 0.
Parallel Claude Code agents in isolated worktrees stacked the queue into PR chains across 4 repos; serial merge gates decided what merged.
- Claude Code
- GitHub Actions
- pytest
04 · Freya · Tech Lead
A writing partner that never loses the plot
Before: 7 agents. After: 1 story graph.
Co-architected an AI writing partner around an "hourglass" model (many input channels, one facts-only knowledge base, prose generation) and hardened its AI layer for production.
- FastAPI
- pgvector
- Yjs / CRDT
Open source
Code you can actually read
Client work stays under NDA. These are public, small on purpose, and each one is about a single idea.
- rag-bm25-searchAnswers with citations, or not at allQuestion answering over internal documents that cites every claim and returns the closest sources instead of an apology when the evidence is weak.Retrieval 6/6 · answering 1/6 → 5/6 — measured, not claimed
- Python
- BM25
- Eval gate
- aws-serverless-fastapi ⇄ ecs-fargateSame API, two runtimesOne URL shortener, one domain layer, deployed twice: Lambda with DynamoDB and Fargate with PostgreSQL. Reading them side by side isolates what the runtime changes.Terraform · Serverless · one shared core
- FastAPI
- AWS
- Terraform
- ai-agentsThe agent setup I work with dailyClaude Code subagents, slash commands and permission settings, versioned and portable across projects — the tooling behind the Reeve merge campaign, generalised.Claude Code today · Cursor and Copilot planned
- Claude Code
- MCP
- Shell
Experience
Eight years, one directionEight years, three continents, one direction
From PHP back offices in Lima to leading AI products for US startups — always on distributed teams across the US, EU and LATAM.
- 2024 — nowAI EngineerIntercorp · Innova LATAMProduction LLM system generating curriculum-aligned lesson plans for 3 countries.Peru · remote
- 2025 — 2026Technical Lead, AI ProductsMindFortressCode review and release gating for Freya and Reeve; the team’s AI-assisted delivery system.USA · remote
- 2022 — 2024Backend & AI EngineerWordboxLLM content pipelines for 500K+ learners; Python → Go rewrite, −40% latency.MX / UK · remote
- 2023 — 2024Backend DeveloperAccelOneVideo platform backend: large-file pipelines on SQS, Elasticsearch search.USA · remote
- 2021 — 2022Backend Developer, freelanceIndra · Pretty TechnicalAPIs serving ML models at 95% coverage; Go services for fintech and genomics.PE / NL
- 2020 — 2022Backend DeveloperRed Científica PeruanaInternationalised a hosting platform serving 11,000+ businesses.Lima
- 2017 — 2020Software EngineerCodriseFrom L3 support to shipping Laravel and CodeIgniter APIs.Lima
- 2017 — 2022Backend & Software EngineerIndra · RCP · Codrise
How I work
Three rules I don’t break
- Evals before vibesA prompt change ships when a fixture set says it is better — not when it reads better.
- No threshold is a guessEvery limit in a system I own has a measurement next to it, and a note on what it costs.
- Cost is a featureTokens, latency and dollars per request are on the dashboard from day one, and reported to whoever pays.
Let’s build something
that holds up.
Open to AI Engineer, Senior Backend and Tech Lead roles across the EU.
Open to AI Engineer, Senior Backend and Tech Lead roles across the EU — remote, or on-site in Milan.