Projects

perfettoagent

LLM agent for Android performance regressions

Takes two Perfetto traces and a git commit range and returns a cited diagnosis: the regressed metric, the size of the regression, the culprit commit, and the trace rows that prove it. A deterministic verifier re-runs every citation before the report is written.

On planted regressions across 20 cases in two Android apps: 93% detection, 100% culprit attribution, 96% citation validity, at USD 0.044 per trace.

GitHub ↗Read the write-up →

pareto-eval

Multi-axis evaluation for LLM agents

Reports agent quality, USD cost, and p95 latency together as a Pareto frontier rather than one score, applying the rate-distortion curve from video compression to model and prompt selection. A Cohen's kappa gate keeps an uncalibrated LLM judge from contributing a quality score.

GitHub ↗Read the write-up →

SuperPlayer

Production playback layer on AndroidX Media3 and AVFoundation

Implements the Media3 Player interface on Android, and wraps AVPlayer on iOS and tvOS, so existing PlayerView, MediaSession, Compose and AVPlayerViewController surfaces take it unchanged. Fourteen modules behind one seam each, QoE telemetry to CTA-2066, and 18 architecture decision records.

GitHub ↗