At Manifold Research Group, I contribute to MultiNet v2.0, an open-source benchmark for AI agents. I added full-context history and a per-turn step budget to the agent runner, and designed prompt-caching support for Anthropic, Kimi, and OpenAI models.
On a project reproducing on-policy distillation, I built a GSM8K evaluation harness and compared the student model with its teacher. I’m running the work on a Colab TPU v5e chip with Tunix.
I built a GPT-style language model from scratch in PyTorch, then implemented PPO, GRPO, and DPO for an RLHF alignment study. The experiments revealed reward-margin explosion in DPO and group collapse in GRPO; targeted tuning reversed the ranking and gave GRPO +3.43 average reward.
My other projects include a modular VLM evaluation pipeline for Qwen3-VL-2B-Instruct and an end-to-end cuisine classifier deployed through a Flask app. I’m an early-career machine learning engineer focused on LLM post-training, evaluation, and end-to-end ML pipelines.

