AI Scientist/Researcher

I am Ambuje, an AI Scientist with 5+ years of experience. If I need to describe my superpower, that will be Reading, implementing research papers. This has always helped me to develop SOTA AI solutions

Would list down some of my work, what I have recently worked on

1. Fine-tuning LLMs (upto 70B)(SFT, RL)
2. When to use LoRA, Tiny-LoRA and RS-LoRA (For domain-specific, we should use RS-LoRA. We would need to understand the mathematics to know the difference between LoRA and RS-LoRA(the difference is just dividing by root(r)))
3. Agent Memory - How we can efficiently manage the state
4. When to use Single vs multi-agent architecture (backed by research, not by gut :) )
5. Solved LLM Latency by smartly caching the planning stage in Agents
6. Build Voice Agents (our TTFT latency was 1.2 sec, best in business for 70B model, for 8B it was less than 100ms)
7. Build a knowledge graph for Agent memory as well as RAG
8. Continuous pre-training of encoder-based models for domain-finetuning. This also included a model fusion to avoid catastrophic forgetting(trained using contrastive loss)
9. Implemented simple prompt engineering ways to make LLM safe by prompt injection attack
10. Agent Evaluation:- Apart from academic benchmark(P, R, F1), developed a product KPI for business in terms of goal completion and progress rate using a self generated worflows graph
11. Built a data generation pipeline for training LLMs(workflow graph-oriented)
12. Agent orchestration, built a framework from scratch in which you can configure your agents(Can use any architecture:- Single agent, swarm agents,supervisor-worker agent, constellation architecture)

Research

1. Learning RL from scratch
2. Know how to make a SLM from scratch (MOE architecture)