available for work

I train custom models
& harness AI agents.

Hi, I'm Aviraj Khare — a decade of shipping software: platform engineering at Gojek, then AI infra and developer tools. Now I train custom models on Apple Silicon and build the harnesses that make AI agents actually work.

aviraj@localhost — zsh
$ whoami
engineer · ex-Gojek platform eng · ETHIndia '22 winner
$ ls ~/current
troy/  superdupermemory/  yoyo/  wasm_terminal/
$ cat status.txt
open to interesting work → cal.com/aviraj/coffee-chat
$

Troy fine-tunes LLMs on your Mac. One YAML file becomes a LoRA training run on Apple Silicon — no notebooks, no boilerplate, no GPU rental.

  • SFT, DPO, ORPO and vision fine-tuning, built on MLX
  • Auto-detects your dataset format — alpaca, sharegpt, chat, completions, preference pairs
  • Chat with, evaluate, and serve the tuned model locally via an OpenAI-compatible API
  • Export to GGUF / Ollama, or push straight to Hugging Face
troy — zsh
$ brew install avirajkhare00/troy/troy
$ troy init && troy train
training… loss 2.41 → 0.87 · adapter saved
$ troy chat -p "hello"
your model, tuned, running on your Mac.
$
April 2, 2026AI InfraGitHub Actions is all you need.January 27, 2026Steer the LLMSteer the LLM, do not let LLM steer youJanuary 24, 2026How I code in 2026My observations on myself while I write codeJanuary 18, 2026Securing Third-Party API Keys Across Multi-Cloud InfrastructureCloud AgnosticFebruary 20, 2026Fun With Stock Market Data - Intraday EditionTwo Beautiful Ideas. Both Killed by Math.August 26, 2025Zig vs Python on Fibonacci>-
all 18 posts