Plate II · Selected work
Five constellations
Each project is charted as a figure in the sky. The case studies carry the evidence: problem, approach, result, with real numbers wherever there are any.
-
sprk-ai autocategorization
Libra
Production transaction categorization at SPRK — a rule-first cascade with one grounded, abstaining LLM call.
-
SignalML
Lyra
An arXiv ingestion + LLM-classification pipeline on Supabase — the data layer of a future reader.
-
metrics-server-dashboard
Horologium
A self-hosted Mission Control — Go WebSocket fan-out at 1 Hz, memory bounded at every layer.
-
Globe visualization
Armillary Sphere
A config-driven 3D globe instrument — seven render strategies, runtime GLSL, live re-mapping.
Plate III · The observer
About
I'm an AI software engineer in Barcelona, building production ML systems by day and charting personal projects — from a hand-rolled VPN to real-time observability — by night.
This site maps that work as a sky: each project a constellation, drawn as the case studies are written.
Observer's log
- Station
- Barcelona, Spain
- 41.4° N · 2.2° E
- Focus
- LLM systems in production, MLOps, and the measurement discipline that keeps them honest
- Degree
- B.Sc. Computer Engineering
- Expected Dec 2026
- Instruments
-
- Python
- Go
- TypeScript
- AWS
- Postgres
Plate IV · Field notes
Writing
-
Personal ·
Something That Matters to MeIf no one needed anything from me—and no one was there at all—would I still have something I wanted to do?
-
Technical ·
Frozen Models Can Drift from Their Dataset HashesWhat a stored content hash protects in Pydantic v2, where shallow freezing falls short, and the measured cost of hashing a dataset twice.
-
Technical ·
Building an Event-Driven AI MonorepoWhy treating AI microservices as async data pipelines — not REST call chains — solves the scaling problems monoliths can't
Plate V · Correspondence
The observatory is open
For roles, collaborations, or a good question — write.