Hello!
I am Logesh Kumar Umapathi , a Machine learning Research Engineer at Blackbox.ai. My work focuses on building agentic systems and optimizing LLMs for high performance serving.I have also worked on Coding agents that achieved SoTA on coding benchmarks like SWE-Bench (verified and lite) and Nextjs Agents Eval.
Outside of work,I am building an open source Video Agent - Ambient . Which won 2 categories and placed runner-up in a third at the ECCV Wearable AI Challenge 2026 - orgnaized by Meta.
Previously, I was a Lead Maching learning Engineer at Saama Technologies. Leading machine learning efforts for a product focused on accelerating clinical trials and time to market of drugs.
I have been part of notable Code LLM research including Starcoder , Santacoder and Bigcode Evaluation harness.
I was recognized as a Top 7 GenAI Scientist at the AV Luminary Awards 2025.
When I’m not in front of a computer screen , I would love to speak at ML events. I also love reading books and photography.
My socials: 
Technologies
🤖 Machine Learning: Transformers, PyTorch, Scikit-Learn, Weights & Bias, Deepspeed, TensorRT , Sglang , vLLM , Text generation inference.
☁️ Cloud: AWS, GCP, Docker, Airflow.
Publications:
-
Ambient @ EgoLongQA 2026: Distilling Long-Video perception into a Sub-2B Model - ECCV 2026 [Code]
-
Ambient @ EgoProactive 2026 : Proactive Egocentric Assistance with Visually Grounded Supervision - ECCV 2026 [Code]
-
Med-halt: Medical domain hallucination test for large language models - CoNLL 2023 Code , Website
-
Starcoder: may the source be with you! - TMLR 2023
-
MedMCQA - A large-scale multi-subject multi-choice dataset for medical domain question answering - ACL 2022
-
Santacoder:Don’t reach for the stars. - ICLR 2022
Notable Talks:
-
Building a Video Understanding Agent: Search, Citations & Reasoning - Analytics Vidhya’s - DataHack summit 2026
-
From Language to Robotics: Practical Lessons Bridging LLMs, RL, and AI - Analytics Vidhya’s - DataHack summit 2025
-
Taming the Large language models – Efficient inference of Multi-billion parameter models - NLP Summit 2023. [Presentation]
-
Bespoke LLMs : Building and Scaling customized large language models - Analytics Vidhya’s - DataHack summit 2023 [code]
-
Unlocking reasoning and planning abilities in Large language models - CONF42 2023
-
Decoding state-of-the-art NLP models at DataHack Summit 2019 - Analytics Vidhya’s - DataHack summit 2019
-
Get your feet wet with ML - Google developer group CBE DevFest 2019. [code]
Open source projects:
-
Ambient : A video agent for any video in and action out. #VideoUnderstanding #Long-formVideo #ECCV2026
-
GUI-Decisions : Decision models for GUI grounding. 3x faster than existing methods. [Blog]
-
Bigcode Evaluation harness : A framework for the evaluation of autoregressive code generation language models. #pytorch #CodeGeneration #LLM
-
Mutate: A library to synthesize text datasets using Large Language Models (LLM) #pytorch #transformers #LLM
-
Keras Scaffolding - A scaffolding for keras and tensorflow with some callbacks , metrics and logs inbuilt.
