ML engineer · IIT Madras · New York · [email protected] · LinkedIn
At Manipal Hospitals I was the first hire on a $33 million government-funded project to detect coronary artery disease with AI. I built a team of 8 and took the project from proof of concept to production. The cardiovascular risk work was published in The Royal College of Radiologists Open, and I established the hospital's AI Center of Excellence. At Subex I then owned the data stack behind a production analytics product handling 180M+ records a day.
I studied computer science and data science at IIT Madras, where my research on hospital pathogens became a publication, and earned my MS at Binghamton on a full-tuition merit scholarship. I also have a first-author paper accepted at HICSS 2027 and a provisional patent in AI for biotechnology.
In open source, I have merged fixes in Microsoft's ONNX Runtime and Apple's MLX, including numeric bugs in ARM64 kernels, and I published quantcost, a tool that measures what quantization really costs on real hardware.
My fixes ship in Microsoft's ONNX Runtime 1.31 and Apple's MLX 0.32.4.
1,024 wrong fp16 outputs in the ARM SVE kernels, fixed by emulating all 16 vector lengths. #33040
32,447 kernel tests CI compiled but never ran. They now gate every PR. #33052
96/96 integer cases where logcumsumexp silently returned a cumulative max. #4625
1 bit of state fixed vmap(irfft) on odd lengths, a limitation an earlier PR left open. #4637
55x faster CoreML quantizer. The Meta reviewer built on it to re-enable skipped int8 tests. #23323
1.5–2x faster int8 decode: ONNX Runtime re-dequantized tied lm_head weights on every token. Found with quantcost, fixed upstream, under review. onnxruntime #33800
forgot names a truly missing file in 70% of the commits it flags, vs 42% for a baseline.
quantcost found int8 runs at 0.47x fp32 on Apple M4, and traced why.
ML risk prediction for cardiovascular disease · RCR Open 2025 · first author
Pathogen persistence in hospital microbiomes · medRxiv 2025