Projects

Flagship builds, in the open. Each ships with architecture notes, code, measurements, failure analysis, and a decision record — not just a README. The status labels are honest: nothing here claims to be more finished than it is.

In progress

Agent Harness Reference Architecture built

The system argued for in the harness essay, turned into working code at github.com/shashforge/agent-harness: the executor state machine, tool contracts, failure taxonomy, append-only trace with crash-resume and replay, golden-trace regression tests, and an LLM planner behind the same callable — each pinned by behavioral tests, plus context compaction that shrinks the model’s view without ever touching the record. Next: the first live-model trace, published unedited; the run kit already sits in the repo. Decisions get recorded as ADRs, starting with ADR-001.

Edge AI Runtime Benchmark designed

One representative model, measured properly across ExecuTorch, LiteRT, ONNX Runtime, and llama.cpp on Android hardware: latency, throughput, memory, binary size, FP16/INT8/INT4 quantization, and accuracy impact across CPU, GPU, and NPU. The measurement protocol is published and frozen before any numbers exist; reproducible scripts ship with the results.

Queued

Production AI Gateway

A provider-neutral inference gateway: model routing, fallback, caching, retries and timeouts, cost attribution, tracing, SLOs, and chaos testing against provider failure.

More on GitHub. If one of these overlaps with a problem you’re hiring for, email me.