4 Aug 2026 · 5 min read
Drop the Bloat
Everything you hand a coding agent, it has to carry. Mine was hauling 35K tokens of instructions before I typed a word. A field note on cutting that to 13K, and on knowing when context stops earning its place.
Lorenzo Germini
AI Product Engineer · Cuneo, Italy, CET
CV →
I build AI products end-to-end and write about what actually works.
Right now that's the AI engine behind Complaion, a compliance platform for ISO certifications: LLM agents doing real compliance work in the background, and the evals that show when they can't be trusted. Before that: pharma manufacturing, digital health, and a GenAI startup.
Start with the writing ↓Writing
Field notes, mostly from building with AI, published on germinai. No definitive guides. Things I tried, what broke, and what earned its place.
About
I'm an AI product engineer: I ship AI products end-to-end, from the models and infrastructure to the interface people actually touch, and stay on the hook for whether the thing helps anyone. Full stack in the literal sense: backend, frontend, and the product calls in between.
These days that's the AI engine behind Complaion, a compliance platform for European SMEs getting and keeping ISO certifications. The interesting problem is making the agents in the background reliable enough to trust with real compliance work, and knowing when they're not. Before that: AI for pharma manufacturing at GSK, AI R&D in digital health, and a GenAI EdTech startup I built as founding engineer. The through-line hasn't changed: technical depth only matters when it meets a real user problem.
Outside work: sports, side projects, and rabbit holes on tech entrepreneurship and biohacking.
Work
May 2025 – Present
AI Engineer · Complaion
The AI engine behind the product: LLM agents doing compliance work over ISO documentation in the background, and the eval harness that shows when they can't be trusted.
Nov 2024 – Apr 2025
Founding AI Engineer · Stealth GenAI Startup
Backend and GenAI API layer from zero on AWS; a RAG stack with custom parsing for handwritten input, plus the tracing and evals to tune retrieval.
Sep 2023 – Oct 2024
AI Engineer · GSK
Time-series forecasting and anomaly detection on plant sensor data, semantic NLP for fault detection, and a water-rejection strategy projected at 10,000 m³/year.
2021 – 2023
- A Stable Diffusion interior-design MVP, killed on the evidence
- Biomedical-literature summarisation at Burgeon Labs
- Fill-parameter ML at Roche
- EPFL MSc, Life Sciences Engineering
Projects
L'Oracolo della Ghigliottina
AI game companion for La Ghigliottina, the cult final round of Italy's most-watched quiz show. Send the five clue words, typed or snapped from the TV, and it finds the one that connects them. 100% accuracy so far, with rate-limit handling because Italians take this game seriously.
Side Project
Live
Systems
agent systems · eval harnesses & tracing · end-to-end shipping, coding agents in the loop