Execution intelligence, in real sessions.

Problem

An IMU tilt feature on a BLE device. Timing and power normally need real silicon — a JTAG probe, a power analyzer, the board on the desk.

Prediction

LOCI predicted per-function timing to nanosecond precision from the firmware binary, and flagged a libm sqrtf software loop on the hot path.

Outcome

A 77× fix on the hot path. ~4–8 hours of bench work replaced by seconds. The board never left the drawer.

Problem

A change encrypted session tokens with AES-256-GCM by creating a new EVP cipher context on every request.

Prediction

LOCI predicted p99 latency 18× over budget — per-request alloc/free on the hot path, throughput regressing — before a single test ran.

Outcome

The agent pooled the context and reused it per call. The throughput regression never reached production.

Problem

An agent raised GEMM register tiling on SM_90 to reuse more data — a change that looks faster in source.

Prediction

LOCI predicted register pressure over budget (132 > 128 regs) — ptxas would spill — read from PTX / SASS, before launch.

Outcome

Re-tiled to single-buffered K. Within budget, 1.32× faster — no spill, no slow inference.