Rahul Kumar
ML Engineer @ Google Core ML (Keras)
Hi, I'm Rahul. I make frontier GenAI models run on the devices in people's pockets, and I contribute the open-source plumbing that gets them there. If it involves Gemma, Llama, or quantized transformers on NPUs, it probably crosses my desk.
⚡ Open-Source Impact
50+ pull requests merged into Google's Keras ecosystem — LiteRT export for the PyTorch backend, Llama 3.1 in KerasHub, torch-backend performance, and the keras.io guide that explains the first one. I don't keep the tally on this page; GitHub does, and GitHub is never out of date.
🏗 Edge Deployment Architecture
How frontier models travel from research to the NPU in your pocket — via Keras and LiteRT.
Keras 3 abstracts the training backend; LiteRT-LM export produces hardware-optimised prefill + decode TFLite graphs. Gemma 3 270M has been validated end-to-end on Android.
💼 Experience
- Shipped LiteRT export for the PyTorch backend in Keras 3, enabling on-device deployment of PyTorch-trained models via TFLite / LiteRT (keras#22758).
- Built LiteRT-LM export for KerasHub with prefill / decode signatures — validated Gemma 3 270M on Android, end-to-end.
- Implemented Keras-native Llama 3.1 and Multimodal Gemma 3; fixed 4/8-bit quantization instabilities across backends.
- Wrote the official Keras guide to LiteRT export — live on keras.io, under my name.
- Migrated Adbrain ad-recommendation model to Keras 3; TF 2.20 / Python 3.13 / Orbax compatibility fixes.
- 50+ PRs merged across keras, keras-hub and keras-io — serialization and quantization fixes, torch-backend performance, cross-backend compatibility (the list is public).
- Applied Quantization-Aware Training (QAT) on Snapdragon mobile SoCs, reducing memory bandwidth 75% while retaining model quality.
- Redesigned Magic Keeper generative inpainting pipeline for the Snapdragon Summit demo — NPU and DSP accelerated.
- Built voice-controlled Camera Copilot demo for on-device AI showcase at client briefings.
- Delivered hardware-accelerated Android reference apps demonstrating GPU, DSP, and NPU delegation for benchmarking.
- Knox Capture: End-to-end damaged-barcode scanning pipeline — 10M+ synthetic samples via geometric/probabilistic transforms; accuracy 70% → 90% with YOLO, Mask R-CNN and U-Net.
- SRIN-Satyapan: Anti-cheat proctored exam platform — 2,000+ concurrent users, 10,000+ registered faces, real-time face-recognition auth on Android + web portal. MD Appreciation Award.
- Face recognition optimisation: Embedding search O(n) → O(log n); lookup time 1,200 ms → 87 ms on Android tablets serving 10k+ users.
- Drop detection: Replaced deep learning with statistical feature engineering — 99% parameter reduction, 95% power reduction, accuracy drop <2%.
- Trained CycleGAN for generative-AI use cases in Samsung's software competency programme; converted to full-time R&D Engineer in July.
🚀 Projects
On-device deployment infrastructure for frontier models (Gemma, Llama, Flux) via Google's Keras 3 ecosystem. Covers PyTorch-backend export, LiteRT-LM prefill/decode signatures, and the official keras.io documentation.
JAX-native PPO and DQN agents trading across 746 NSE instruments with 3.5+ years of OHLCV data. 128 vectorised environments for parallel rollout, full portfolio simulation with risk controls.
A 270M-parameter Gemma generating tokens on a phone, end-to-end: KerasHub → LiteRT-LM export with separate prefill and decode signatures, then a minimal Android app that runs it on device.
🛠 Skills
🏆 Achievements
🎓 Education
👋 Let's Connect
Open to collaborations, interesting problems, and good coffee chats.