Keras on PyTorch: compiled generation 11.6x faster
Traced why Keras on PyTorch ran about 9x slower than plain PyTorch, and merged the fixes upstream, one root cause each.
Case study · 2 minMerged →
Experience
Where the work was done
Each role, then the work from it. Cards are the projects worth opening; the rest is listed below them.
- Feb 2025 – now · India
Software Engineer, Machine Learning
Google
Now on XProf, the open-source profiling tool for machine learning. Before that, on Keras and KerasHub.
One call from a Keras model to a phone: LiteRT export
Built the LiteRT export path in Keras and KerasHub: a trained model to a phone-ready file in one call, on TensorFlow and PyTorch.
Case study · 1 minMerged →End-to-end UI tests for XProf
Wrote the end-to-end UI test suite for XProf, the open-source ML profiler, with 107 browser tests and a side-by-side check of two builds.
SummaryShipped →Llama 3.1 in KerasHub, and Gemma 3 fixes
Implemented Llama 3.1 in KerasHub and fixed Gemma 3 bugs, including a preprocessor crash across Gemma 3, 3n and 4.
SummaryMerged →A benchmarking system for Keras and KerasHub
Built the benchmarking system for all Keras and KerasHub models: automated weekly runs across GPU and TPU, the TensorFlow, JAX and PyTorch backends, presets and batch sizes, to catch performance regressions.
Short entryBuilt →- Keras correctness fixesJun 2025 – Oct 2026
Fixed TPU attention, N-D attention inputs, nested weight saving, masks, call-context leaks and ONNX export of models over 1.5 GB. · Merged
- Model-parallel layout maps for KerasHub LLMsJul 2026
Designed model-parallel layout maps for KerasHub LLMs, from Gemma and Qwen to Llama and Mixtral; proposed upstream, not merged. · Proposed, not merged
- A Keras 3 migration
Led the migration of a major internal customer from legacy Keras to Keras 3, with backward compatibility and performance parity. · Built
- Making KerasHub models exportable
Fixed the layers that kept KerasHub models such as ESM, PARSeq and Flux from exporting to LiteRT. · Merged
- LiteRT-LM export for KerasHub LLMs
Built one call from a KerasHub LLM to a LiteRT-LM bundle with tokenizer and prefill and decode signatures; proposed upstream, not merged. · Proposed, not merged
- Keras correctness fixesJun 2025 – Oct 2026
- Sep 2024 – Jan 2025 · Hyderabad, India
ML Engineer, Computer Vision
Qualcomm
On-device AI for Snapdragon phones.
Generative inpainting on a phone's NPU and DSP
Redesigned a generative inpainting feature, object removal and image completion, so its models fit NPU and DSP latency budgets for the Snapdragon Summit demo.
SummaryDemo →3x faster models on the Snapdragon NPU and DSP
Made deep-learning models run 3x faster on the Snapdragon NPU and DSP, and cut memory bandwidth 75% with quantization-aware training.
SummaryBuilt →- Camera ML pipelines in C++Sep 2024 – Jan 2025
Detection, tracking and segmentation pipelines in C++ with INT8 quantization and pruning, and ML inside the native camera stack for a voice-controlled camera assistant. · Built
- Camera ML pipelines in C++Sep 2024 – Jan 2025
- Jul 2022 – Aug 2024 · Noida, India
Research and Development Engineer
Samsung R&D Institute India
Computer vision, biometrics and on-device AI, plus the apps, portals and servers around them.
Reading damaged barcodes on phones
Read 90%+ of damaged barcodes correctly on its damaged test set, where a commercial scanning SDK read 70%, with models trained on 10M+ synthetic images.
Case study · 2 minBuilt →Recognising many faces at once on edge devices
A multi-face recognition engine at 99.2% accuracy and 30+ FPS on edge devices, with face search cut from 1,200 ms to 87 ms.
Case study · 1 minShipped →Phone drop detection in under 22 parameters
Replaced a deep-learning drop detector of millions of parameters with sensor features and a model of under 22, saving over 95% of the power.
Case study · 1 minShipped →Who is holding the phone? Behavioural identification
Told which hand holds a phone, right 90%+ of the time over 30+ seconds, and who holds it, without deep learning.
Case study · 1 minShipped →Car crash detection with automatic SOS
A phone that detects a car crash from its sensors and sends an SOS to chosen contacts.
SummaryBuilt →Telling attention from distraction, with emotion recognition
Detects whether a person at a screen is attentive or distracted, at 87% accuracy; checks across frames cut false alarms by 40%.
SummaryShipped →An AI exam-proctoring platform (Satyapan)
An anti-cheat exam platform used for 2,000+ employees across 100+ tests, with 60% less administrative overhead.
SummaryShipped →- Shadow removal and document clean-up
Prototyped shadow removal and noise reduction for document scanning on phones. · Research
- Faster Exynos NPU inference
Optimized models for the Exynos NPU with TensorFlow Lite, for 2.5x faster inference. · Built
- Shadow removal and document clean-up
- Jan 2022 – Jun 2022 · Noida, India
Intern
Samsung R&D Institute India
Learned Android in Java and Kotlin, and joined full time in July 2022.
- Gaze correction for video callsJan 2022 – Jun 2022
A prototype that shifts the eyes toward the camera in video calls, using facial landmarks and geometric transforms. · Prototype
- Real-time image models on AndroidJan 2022 – Jun 2022
Real-time classification and detection on Android at 25+ FPS with quantized MobileNet and EfficientNet models through NNAPI. · Built
- Generative image modelsJan 2022 – Jun 2022
Trained CycleGAN models in PyTorch for image-to-image translation and avatars, deployed with TFLite, and generative face models with StyleGAN and StarGAN. · Built
- Gaze correction for video callsJan 2022 – Jun 2022
- 2020 – now · alongside the jobs
Independent work
An independent research lab, an LLM served from one GPU, and the tools built along the way.
GLM-5.3-Flash on one GPU
Served a 181 GiB mixture-of-experts LLM from one 96 GiB GPU and took decoding from 10.8 to 15.0 tokens/s shipped, 16.5 at peak.
Case study · 2 minPublished →Keras on an Android phone
An Android app and guide that run KerasHub models on the phone with LiteRT, from one export call to six working screens.
SummaryPublished →RL Alpha Labs
Automated trading on Indian stocks, where a strategy counts only if it beats an index fund after real costs and tax.
Case study · 1 minResearch →Porchlamp
Shows which open-source projects reply to, review and merge a stranger's pull request, so contributors spend their effort where it will land.
SummaryLive →Servedeck
Runs local LLMs on one GPU workstation, switching models with one command behind one OpenAI-compatible endpoint, with a live dashboard.
SummaryLive →
Education, certificates and awards are on their own page.