About

I’m an AI engineer in Miami. I work on large language models and agentic systems. Right now I’m co-founding PromptGuard, where we build safety and threat-protection infrastructure for LLM agents.

Before that I spent a year and a half at Multiverse Computing building a 30B-parameter language model from scratch and co-authoring CompactifAI, a tensor-network compression method that cut model size by around 70 % with little accuracy loss. Earlier I built the voice search that ran for Flipkart’s 200M+ daily users.

Most of what I write here comes out of that work, usually the parts that took longer to figure out than they should have.

Work

Co-founder & AI engineer, PromptGuard 2025–

A multi-agent safety engine for auditable collaboration between enterprise LLMs. Grew out of freelance work on agentic automation prototypes that turned into the core of the product.
AI consultant, PwC 2025–

Part-time with the global AI innovation group: agentic automation, LLM deployment strategy, and dragging MLOps practice forward across departments.
Applied AI scientist, Radiant Science 2025

Contract. HIPAA-compliant agentic booking and multimodal systems for healthcare operations, with human-in-the-loop evaluation and privacy-first data pipelines.
Senior ML scientist, Multiverse Computing 2024–25

Built a 30B-parameter enterprise LLM from scratch, with the training and evaluation pipelines around it. Co-authored the quantum-inspired tensor-network compression work, and cut evaluation iteration time by about a third by automating it.
AI researcher, TheoremOne 2023–24

Frontier prototypes for client and internal R&D. Built CodeMe, a VS Code assistant that could ingest a whole codebase before GPT-4 made that ordinary, and led a team on RAG pipelines that beat TableQA baselines for structured QA.
Applied AI scientist, Flipkart 2021–22

Voice search in production for 200M+ daily users, improving transcription accuracy by 40 %.

Earlier: ML research intern at Optum (2018–19), building clinical diagnostic models and the first traceable pipelines around them; freelance full-stack and data science work before that.

Open source

quantum_hybrid_model Benchmarking framework for quantum-classical hybrid models.
mmrag Multimodal RAG question-answering system.
lora-finetune LoRA fine-tuning platform.
llm_training One CLI for the usual LLM training pipelines.
rl-token-compression Token compression learned with RL.
autorank-llm Automated LLM benchmarking and ranking.
ai_command_automation AI-driven command automation for macOS.

The rest is on GitHub.

Publications

Tomut, A., Sarkar, A., et al. “CompactifAI: Extreme Compression of Large Language Models using Quantum-Inspired Tensor Networks.” arXiv:2401.14109, 2024.
Sarkar, A., Sahoo, A. K., Sah, S., Pradhan, C. “LSTMSA: Stock Market Prediction Using LSTM and Sentiment Analysis.” IEEE ICCSEA, 2020.

Education

M.Tech, AI & computer science, at IIT Ropar 2019–21

Thesis: DocYOLO, optimising object detection for document layout extraction.
B.Tech, computer science, at KIIT University 2015–19

All India Rank 498 in GATE 2019: top 0.5 % nationally in computer science.

Elsewhere

Reach me at abhijoy.sar@gmail.com, or on any of the links in the footer. If you want to work together, here’s how that usually goes.