AI researcher and Principal AI Engineer. Indian, based in Dubai. My work sits where post-training, agents, evaluation, and inference meet: training open-weight models to do specific jobs well, building the systems that run them in production, and measuring whether they actually got better. I contribute upstream to the projects that stack depends on.
12 merged upstream across 4 repos. Updated 2026-09-30.
| Repo | Merged | Summary |
|---|---|---|
| unslothai/unsloth | 6 merged | These changes stop the chat and studio UI from losing state: queued or stopped prompts, saved context, compare-pane settings, and sandbox files survive interruptions, model switches, and snapshot paths, and the CLI API key is reused across runs. |
| google/adk-python | 3 merged | Keeps LlmAgent config backward compatible, fixes path handling for jj and hg users on Windows, and surfaces MCP grounding metadata on events so agents can trace where tool responses came from. |
| twentyhq/twenty | 2 merged | Fixes edge cases around the percent sign: percentage field values entered with a trailing "%" now save correctly, and array containsIlike filters treat "%" as a SQL wildcard as intended. |
| e2b-dev/E2B | 1 merged | Sandbox creation now checks the options you pass before it asks for an API key, so a bad config fails fast with a clear error instead of an unrelated auth complaint. |
Fine-tuned and shipped Qwen3.6-27B with QLoRA/SFT, then on-policy context distillation from a 24.8K-token expert-rule corpus. Merged adapters into a standalone bf16 checkpoint and served it on H100s with vLLM at 73K context.
Built the self-improving loop around it: production generations become critic-ranked Best-of-N preference pairs and broken-to-repaired training examples for DPO and RLVR/GRPO, with deterministic verifiable rewards and frozen held-out gates deciding whether a candidate model is promoted or rejected. On a locked 30-task by 3-seed harness, rule distillation cut mean defects 54% (0.79 to 0.36 per output) and raised strict zero-defect passes from 17% to 53%.
DeepField: an independent research program on whether the early signatures of significant events can be detected ahead of consensus.
Earlier, a CNN for object amplification in IBM Watson PowerAI Vision.
CiaraAI. Stateful agent platform running autonomous multi-step workflows across CRM, lead routing, scheduling, payments, voice, WhatsApp, and chat. 200K+ interactions, 97% automation, sub-2s responses, about 70% lower cost than the stack it replaced.
Claws. Visual multi-agent orchestration on the OpenClaw runtime for manager-worker-reviewer teams. Workflow definitions compile to executable agent configs with concurrent execution, tool integrations, runtime monitoring, and failure recovery.
Raven. Open-source AI meeting copilot: system audio plus mic capture, WebRTC AEC3 echo cancellation, real-time transcription, live assistance. 400+ stars.
Evaluation and observability. A 559-rule learning-science evaluation framework catching about 92% of defects before human QC, and root-cause instrumentation across 500K+ production AI failures, turned into regression signals.
Stack. Python, TypeScript, PyTorch, Unsloth, vLLM, LangGraph, PostgreSQL, Kubernetes, AWS, GCP. Claude Certified Architect (Foundations).