AI Engineer specializing in LLM alignment, agent systems, and deployment.
I build and scale production AI systems end-to-end. Previously, I was the Lead AI Engineer at Fern-Bio, where I led a cross-functional team of 4 engineers and built our complete model alignment, training, and multi-agent systems serving thousands of daily production requests. I specialize in fine-tuning open models (9B to 120B), reinforcement learning alignment (GRPO, DPO, PPO), low-latency inference, and bridging AI internals with outstanding developer experiences.
Led the development of the entire AI platform as Lead AI Engineer: built instruction datasets, aligned domain-specific models via GRPO/DPO, orchestrated multi-agent reasoning networks, and ran low-latency production APIs serving 3,000+ requests/day at 99.99% uptime.
Optimized inference for training-free video matting with pretrained video diffusion models, enabling high-quality segment masks at high framerates.
Ported and optimized RF-DETR for Apple Silicon, enabling low-latency, real-time object detection and segmentation workflows on macOS.
Regularised linear model experiments and ML fundamentals work, part of my older but still useful learning and implementation base.
Agent harness experiments for composing multi-step AI workflows.
Built a lightweight retrieval-augmented generation pipeline with hybrid keyword-semantic search and custom context compression for fast Q&A.
Led and coordinated a cross-functional team of 4 engineers responsible for model training infrastructure, inference optimization, and synthetic data generation. Built the company's complete model alignment, training, and multi-agent platforms from scratch, serving 3,000+ daily production requests at 99.99% uptime. Managed enterprise technical integrations and scope with partners including Qualcomm, Marvell, and UC Berkeley.
Built a video understanding model using keyframe captioning and temporal aggregation, processing hour-long footage on a single NVIDIA T4 GPU. Designed a high-speed audio pipeline with speaker diarization and transcription, and fine-tuned Llama 3.1 70B on 4x A40 GPUs to generate synthetic datasets from connected enterprise sources.
Rebuilt the prospect research pipeline with ML models and SQL automation, saving $40K/year in manual overhead. Fine-tuned BERT-based classification models to improve outreach open rates by 16%, and co-developed a GPT-3.5 outreach tool that increased client delivery throughput by 7%.
Built and deployed custom LLM, RAG, and multi-agent systems for startups (including Tensorlake) and mid-market tech companies, tailoring API integrations and custom embedding/retrieval systems to cut query latency by 30%.
Invited speaker at 15+ technical workshops covering LLM fine-tuning, reinforcement learning, and agentic architectures, helping developers adopt frontier AI.
Active developer focused on deep learning optimizations and porting key research papers and model weights to run with custom optimizations on M-series Apple Silicon platforms (llama.cpp, JAX, vLLM).
Served as an expert judge for 3 AI-focused hackathons and coding competitions. Won Smart India Hackathon 2022, competing against 160,000+ participants from 2,235 institutions.
Peer-reviewed paper on efficient on-device object detection for assistive prosthetic systems.
Python, PyTorch, Transformers, LLM Fine-Tuning, Model Alignment (GRPO, DPO, PPO), Multi-Agent Systems, High-Performance RAG, low-latency APIs, Docker, Git, GCP, Apple Silicon Inference (llama.cpp, JAX), Robotics, Embedded AI.
Hardware is a serious side of my work. I have built from scratch hardware projects ranging from robot arms and self-balancing bots to autonomous warehouse management swarm robots. I am currently exploring the intersection of robotics and AI, with efficient edge inference as my focus.
I am most active on X, which is also the best place to contact me. You can also find my code on GitHub.