$ cat now.md
What I'm doing now
Last updated: · inspired by nownownow
building
- RoiGenius: hardening the revenue attribution platform — webhook reliability, calculator coverage, and cost governance on GCP.
- A trained filter deciding what stays in an agent's context, running in shadow mode against the age-based policy it wants to replace — the numbers are public.
- A prompt-injection guard serving live behind a 200 ms deadline, fail-open by design: if the model is down, the pattern verdict stands.
- Quantitative systems: operating and monitoring automated trading strategies around the clock.
learning
- Post-training: SFT/RLHF data pipelines on self-hosted compute — currently active learning (a queue of uncertain samples mined from production logs) and hard-negative mining.
- Harness engineering: hooks, skills, and adversarial verification for autonomous agents.
- Reading myself wrong in public: the typed judge agrees with its own derived corpus 20% of the time against an 85% bar, and that number stays up until it moves.
where
- Brazil — working remote across timezones.
- Paris alumni heart: EPITA exchange, Spring 2023.