backend systems / applied ai / science

Siddharth Ahire.

I build the plumbing between intent and execution.

I like the parts of software that sit between messy intent and reliable execution: gateways, backend systems, and the unglamorous contracts that stop them from lying. Somewhere along the way I studied Engineering Physics at IIT Madras, which left me interested in what applied AI can help us understand or build in STEM and climate science. Earlier I made perception run 7–8× faster on dual-arm humanoid inspection robots in Tokyo; now I work on persistent, long-horizon agent systems that can hold context, coordinate work, replay decisions, and survive beyond one heroic prompt. Most of the time I am trying to make complicated systems legible, deterministic, and slightly less full of bullshit.

Open source, lately

HRM-Text #5open · PyTorch SDPA fallback for AMD / ROCm Orca #8361merged · resilient serve recipe parsing Orca #8358merged · literal Git paths over SSH relay Colibri #85merged · English user-facing runtime output

Things with my fingerprints on them

01Intent RouterA deterministic TypeScript engine that turns capability, residency, context, and budget constraints into explainable model choices, with exact cost math and replayable traces.02RoutrAn OpenAI-compatible gateway across 45+ providers and 250+ models. I worked on adapters, routing and fallback, circuit breakers, credentials, telemetry, metering, and billing.03AargusA multi-tenant signal agent that reads HN, GitHub, Reddit, and RSS, scores what matters, degrades safely, and drafts actions for a human to approve.