logcat.ai is the AI engineering layer for the operating-system stack. We build agentic systems for Linux and Android devices: we ingest the diagnostic artifacts a device emits (bugreports, logcat, dmesg, modem traces, ramdumps, perfetto, pcap, CAN) and help device makers diagnose, remediate, test, and author features across their OS stack, grounded in their own BSP. Think Datadog and Cursor combined for the OS layer. A VC-backed startup with a small, senior team — you work directly with the founders.

Own the engine that turns every investigation into a compounding asset: the more the platform runs, the sharper it gets. This is applied ML systems, not research. If your goal is training foundation models from scratch and publishing, this is not the seat.

Location: Remote (4h overlap with Pacific or IST; Seattle or Bengaluru on-site preferred). Full-time.

The bar

• 10+ years of relevant engineering experience.

• Production ML systems experience end to end: data pipelines, fine-tuning, eval, deployment.

• Hands-on fine-tuning and distillation with open-weight models (Qwen, Llama-class).

• Eval harness design for correctness-sensitive tasks, including trajectory-level evals for agentic and tool-use systems.

• Inference deployment and scaling (vLLM or equivalent) on managed platforms.

• Proficient in day-to-day work with CLI-based AI coding tools like Claude Code (or an equivalent). It's how the team operates, not a nice-to-have.

What you'll do

• Instrument the product so every investigation, especially human-corrected ones, becomes structured training data.

• Build the eval harness for diagnostic accuracy. Root-cause correctness has to be measured, not eyeballed.

• Distill expensive deep-investigation runs into small, fine-tuned models that hold the quality bar at a fraction of the cost.

• Deploy inference, including on-prem and airgapped SLMs for customers whose logs cannot leave their environment.

Bonus

• Agent or tool-use systems.

• Preference optimization (DPO/GRPO-style) and structured-output / function-calling fine-tunes.

• On-prem or airgapped model deployment.

• Retrieval over large heterogeneous corpora.

• Systems or infra background.

In your application, tell us what you'd own here, and the hardest thing you've shipped in this domain.