Dev Log
Neurova
2k+ users on a SwiftUI attention-budgeting app — real telemetry, real retention, real bug reports.
Code in Place · final project
Built and submitted under Stanford's CS106A — Python from first principles to a shippable artifact.
Public Apple Developer · 5 yrs Swift / Python
Verified Apple Developer with public App Store presence. Five years of compounding reps across Swift (iOS 14–15 device fleets) and Python (research notebooks → Vertex AI pipelines).
Cutting Gemini latency in agentic loops by 47%
Streaming intermediate tool-calls back to the UI while the planner is still cooking. The trick: a tiny optimistic state machine on the client that commits when the server confirms the call signature.
Why my agent kept hallucinating tool args
Turns out: I was passing the schema after the user message. Reversed the order, dropped temp to 0.3, and rolled my own retry-on-validation. Tool fidelity went from 71% → 96% across 1,200 evals.
Productive friction, in code
First MindTrace experiment: insert a 4-second hypothesis prompt before the model responds. Bounce rate barely moved (-2%). 7-day recall jumped 31%. Friction isn't the enemy. Mindless flow is.