Dhruv Nigam

ML Engineer & Researcher

I build systems that build models. Models are the outcomes of well-designed training and feedback loops. My ambition is to build the systems that build and deploy the next generation of AI.

Recently my team and I launched the first live sports AI companion for cricket streams. Speech-to-speech over WebRTC with a diffusion-based avatar at sub-2s latency. 1.5M minutes of watchtime and 1.1M concurrent viewers at peak during IPL 2026. Fine-tuned Llama 8B with GRPO to make it genuinely funny. I gave a talk on scaling this from demo to a million users.

Before that, I trained a 200M-parameter foundation model (LUMOS) on 250M users that replaced 50+ task-specific models overnight. Increased DAU by 3.15% overnight over a base of 13 million. Validated my hypothesis that scaling laws emerge across domains, not just text.

Earlier, I worked as a quant. $1bn/day in client trades at sub-10ms latency, and a stat-arb strategy that returned $4M on $10M capital over two years (Sharpe 1.3). That experience taught me rigor in engineering and modeling.

Interested in LLM post-training and agentic systems. Active contributor to LiveKit Agents (~10K stars) and HuggingFace TRL (~20K stars).