Writing · notes from the workbench
Thinking out loud.
Short essays on what I've learned building AI systems that are fair, robust, and quiet enough to trust in production.
- Apr 12, 20266 min read
Why fairness deltas matter more than accuracy
A model that is 92% accurate overall can still be catastrophically unfair to the 8% it misses. Here is how I learned to stop chasing global accuracy — and start auditing the gap.
FairnessBiasSilentBias - Mar 2, 20265 min read
Notes on building voice-first mental health tools
Voice carries more than words. Building Voice Guardian taught me why prosody is the signal, latency is the UX, and consent is the whole product.
Voice AIMental HealthVoice Guardian - Feb 4, 20267 min read
Agents that hand off work: lessons from DARA v2
Multi-agent systems fail in ways single models do not. Here is what building an autonomous bug-resolution pipeline taught me about handoffs, sandboxing, and knowing when to stop.
AgentsLLMDARA