about
My work sits where a model meets hardware that was not sized for it.
01
how i work
- Start from the constraint. Thermal envelope, memory ceiling, latency budget. The constraint decides the architecture; it is not something to optimize for afterwards.
- Keep the probabilistic part contained. A model is good at reading messy input. It should rarely be the thing making the final decision in a system people depend on.
- Generate the pipeline, do not write it. If the experiment space grows, the pipeline code should not grow with it.
- Measure honestly. Most reported success rates are measured through a gate that quietly changes them — which is what my TMLR paper is about.
02
work with
- languages
- Python · C/C++ · JavaScript
- ml & edge
- PyTorch · NVIDIA TensorRT · LiteRT · LLM quantization · inference optimization · NVIDIA Jetson · Qualcomm 8295p
- cloud & infra
- AWS · GCP (Vertex AI, Cloud Run) · Kubernetes · Docker · Terraform · Apache Airflow
- frameworks
- React.js · Node.js · Express.js · Flask · Django · React Native
- databases
- PostgreSQL · MySQL · MongoDB · Redis
03
research
Below the Reliability Floor: Recovering True Success from Judge-Gated Loops
Transactions on Machine Learning Research (TMLR) · August 2026
J2C certification · selected for the NeurIPS 2026 journal track
04
education
- degree
- B.Tech, Computer Science and Engineering
- school
- SRM Institute of Science and Technology, Chennai, India
- years
- Jul 2021 – Jul 2025
05
credits
- Finalist — Google Cloud Agentic AI Day 2025
Hack2Skill, Bangalore · Jul 2025 - Finalist — HustleX 1.0 Hackathon
IIM Lucknow Enterprise Incubation Center · Aug 2023 - Oracle Cloud Infrastructure 2024 Generative AI Certified Professional
Jul 2024 - Oracle Cloud Infrastructure 2023 Certified Foundations Associate
Jun 2024