Javier Hengda Wu Lin
AI Software Engineer · Barcelona, Spain
Summary
Sole owner of a production ML service at a fintech startup: rebuilt a transaction-categorization system end to end on AWS — architecture, pipeline, evaluation, infrastructure, CI/CD. B.Sc. Computer Engineering (Dec 2026). Focus: LLM systems in production, MLOps, and the measurement discipline that keeps them honest.
Experience
SPRK Software & Tech
Mar 2026 – present · Barcelona
ML Engineer — sole owner of the AI categorization service · case study
- Rebuilt an inherited prototype into a production service: 4-package Python monorepo (~60k LOC, 1:1 test ratio, 1,517 commits) on AWS — ECS Fargate, CDK, RDS PostgreSQL, Redis, Bedrock — with 39 RFCs documenting every architecture decision.
- Designed an 11-step rule-first categorization cascade ending in a single grounded LLM call with the right to abstain — replacing a 3-provider consensus cascade after a measured correlated-errors analysis.
- Built reproducible evaluation: 7,584-row judged corpus and deterministic replay reproducing 1,624 live model verdicts byte-identically at zero cost.
- GDPR-grade engineering: PII encryption with key rotation, HMAC-verified per-user isolation with row-level security in depth, erasure as processing block; hermetic CI with a real-infrastructure merge gate.
Selected projects
- Go
- Linux TUN
- TLS 1.3
- Docker
- A layer-3 VPN in 560 lines of stdlib-only Go: TUN device via raw TUNSETIFF ioctl, length-framed IP over CA-pinned TLS, goroutine relay with first-error-wins semantics.
- Reproducible Docker topology proves the properties: dual tcpdump inner/outer capture, break-and-recover with ~2 s automatic reconnect.
- Go
- WebSockets
- Alpine
- ApexCharts
- Spec-first build: 1 Hz hub fan-out over WebSockets with drop-on-slow-consumer delivery and memory bounded at every layer.
- Embeddable agent SDK with zero non-stdlib imports; adaptive chart density (line → per-core heatmap past 5 cores).
- Supabase
- pg_cron
- OpenAI
- Operations-first design: run-log state machines written before external I/O, per-paper failure isolation, dry-run modes, strict structured outputs double-validated, per-run cached-token cost accounting.
Kofre — personal & social banking backend
2025
- Java 21
- Spring Boot
- GCP
- Terraform
- Team university project: 50+ endpoint REST API (Clean Architecture, DDD); sub-3-minute zero-downtime deploys to Cloud Run; 5 Terraform modules cutting provisioning to under 10 minutes.
Education
University of Barcelona — B.Sc. Computer Engineering
Sep 2022 – Dec 2026
Machine Learning · Computer Vision · Reinforcement Learning · Distributed Systems · Computer Security · Graphics & Data Visualization · Databases
Skills
- Languages
- Python · Go · TypeScript/JavaScript · Java · SQL · C/C++
- AI/ML
- LLM systems in production (Bedrock, OpenAI, structured outputs, grounding & abstention) · evaluation harnesses & deterministic replay · XGBoost · embeddings
- Backend
- FastAPI · Spring Boot · Node.js · REST · gRPC · WebSockets
- Cloud & DevOps
- AWS (ECS Fargate, CDK, RDS, IAM, Bedrock) · GCP (Cloud Run, Cloud Build) · Docker · Terraform · GitHub Actions · GitLab CI
- Data
- PostgreSQL · Redis · Supabase · Firestore · SQLite
- Practices
- TDD · RFC/ADR decision records · hermetic testing · Clean Architecture · CI/CD · observability (Honeycomb, Sentry)
- Spoken
- Spanish (native) · English (fluent) · Catalan (fluent) · Chinese (conversational)
Certifications
- Mathematics for Machine Learning — Imperial College London (2024)
- IBM Data Analyst Professional Certificate (2024)