# pctablet505 (Rahul Kumar)

> Big models, small places, real numbers. The model you need rarely fits where it has to run, and the number you are shown is rarely the one that decides. I work on both: vision and sensing on phones, large models on one GPU, and automated trading after costs and tax.

Every fact on this site comes from one file. Full text: https://pctablet505.github.io/llms-full.txt. Structured: https://pctablet505.github.io/profile.json.

## Performance and inference

- [Keras on PyTorch: compiled generation 11.6x faster](https://pctablet505.github.io/work/keras-pytorch-speed/): Traced why Keras on PyTorch ran about 9x slower than plain PyTorch, and merged the fixes upstream, one root cause each.
- [GLM-5.3-Flash on one GPU](https://pctablet505.github.io/work/glm53-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.
- [One call from a Keras model to a phone: LiteRT export](https://pctablet505.github.io/work/keras-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.
- [End-to-end UI tests for XProf](https://pctablet505.github.io/work/xprof-ui-tests/): 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.
- [Llama 3.1 in KerasHub, and Gemma 3 fixes](https://pctablet505.github.io/work/llama-kerashub/): Implemented Llama 3.1 in KerasHub and fixed Gemma 3 bugs, including a preprocessor crash across Gemma 3, 3n and 4.
- [Keras on an Android phone](https://pctablet505.github.io/work/keras-ondevice-android/): An Android app and guide that run KerasHub models on the phone with LiteRT, from one export call to six working screens.
- [3x faster models on the Snapdragon NPU and DSP](https://pctablet505.github.io/work/snapdragon-speed/): Made deep-learning models run 3x faster on the Snapdragon NPU and DSP, and cut memory bandwidth 75% with quantization-aware training.
- [A benchmarking system for Keras and KerasHub](https://pctablet505.github.io/work/benchmarking-system/): 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.
- [Keras correctness fixes](https://pctablet505.github.io/work/keras-fixes/): Fixed TPU attention, N-D attention inputs, nested weight saving, masks, call-context leaks and ONNX export of models over 1.5 GB.
- [Model-parallel layout maps for KerasHub LLMs](https://pctablet505.github.io/work/layout-maps/): Designed model-parallel layout maps for KerasHub LLMs, from Gemma and Qwen to Llama and Mixtral; proposed upstream, not merged.
- [A Keras 3 migration](https://pctablet505.github.io/work/keras3-migration/): Led the migration of a major internal customer from legacy Keras to Keras 3, with backward compatibility and performance parity.
- [Making KerasHub models exportable](https://pctablet505.github.io/work/kerashub-litert-compat/): Fixed the layers that kept KerasHub models such as ESM, PARSeq and Flux from exporting to LiteRT.
- [Faster Exynos NPU inference](https://pctablet505.github.io/work/exynos-npu/): Optimized models for the Exynos NPU with TensorFlow Lite, for 2.5x faster inference.
- [LiteRT-LM export for KerasHub LLMs](https://pctablet505.github.io/work/litert-lm-export/): Built one call from a KerasHub LLM to a LiteRT-LM bundle with tokenizer and prefill and decode signatures; proposed upstream, not merged.
- [Gemma on Android](https://pctablet505.github.io/work/gemma-android-demos/): Gemma running fully on an Android phone with TFLite, LiteRT and LiteRT-LM.

## Computer vision

- [Reading damaged barcodes on phones](https://pctablet505.github.io/work/damaged-barcodes/): 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.
- [Recognising many faces at once on edge devices](https://pctablet505.github.io/work/face-recognition/): 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.
- [Generative inpainting on a phone's NPU and DSP](https://pctablet505.github.io/work/npu-inpainting/): Redesigned a generative inpainting feature, object removal and image completion, so its models fit NPU and DSP latency budgets for the Snapdragon Summit demo.
- [Telling attention from distraction, with emotion recognition](https://pctablet505.github.io/work/attention-detection/): Detects whether a person at a screen is attentive or distracted, at 87% accuracy; checks across frames cut false alarms by 40%.
- [An AI exam-proctoring platform (Satyapan)](https://pctablet505.github.io/work/exam-proctoring/): An anti-cheat exam platform used for 2,000+ employees across 100+ tests, with 60% less administrative overhead.
- [MILAP: reuniting lost people at mass gatherings](https://pctablet505.github.io/work/milap-design/): An offline-first design for reuniting lost people at the Kumbh Mela, with face and body re-identification, an opt-in gallery and human-verified matches.
- [Object detection tutorial](https://pctablet505.github.io/work/object-detection-tutorial/): A notebook and a synthetic shape dataset that teach box regression and classification from scratch.
- [Camera ML pipelines in C++](https://pctablet505.github.io/work/camera-pipelines/): 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.
- [Shadow-Removal](https://pctablet505.github.io/work/shadow-removal/): Removes shadows from document photos with adaptive thresholding and iterative shading and reflectance decomposition, without deep learning.
- [Gaze correction for video calls](https://pctablet505.github.io/work/gaze-correction/): A prototype that shifts the eyes toward the camera in video calls, using facial landmarks and geometric transforms.
- [Real-time image models on Android](https://pctablet505.github.io/work/on-device-classification/): Real-time classification and detection on Android at 25+ FPS with quantized MobileNet and EfficientNet models through NNAPI.
- [Generative image models](https://pctablet505.github.io/work/generative-image-models/): Trained CycleGAN models in PyTorch for image-to-image translation and avatars, deployed with TFLite, and generative face models with StyleGAN and StarGAN.
- [Shadow removal and document clean-up](https://pctablet505.github.io/work/document-restoration/): Prototyped shadow removal and noise reduction for document scanning on phones.

## Sensor and behaviour ML

- [Phone drop detection in under 22 parameters](https://pctablet505.github.io/work/drop-detection/): 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.
- [Who is holding the phone? Behavioural identification](https://pctablet505.github.io/work/behavioural-identification/): Told which hand holds a phone, right 90%+ of the time over 30+ seconds, and who holds it, without deep learning.
- [Car crash detection with automatic SOS](https://pctablet505.github.io/work/crash-detection/): A phone that detects a car crash from its sensors and sends an SOS to chosen contacts.

## Projects and tools

- [RL Alpha Labs](https://pctablet505.github.io/work/rl-alpha-labs/): Automated trading on Indian stocks, where a strategy counts only if it beats an index fund after real costs and tax.
- [Porchlamp](https://pctablet505.github.io/work/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.
- [Servedeck](https://pctablet505.github.io/work/servedeck/): Runs local LLMs on one GPU workstation, switching models with one command behind one OpenAI-compatible endpoint, with a live dashboard.
- [hostops](https://pctablet505.github.io/work/hostops/): Drops a GPU's power limit the second the UPS reports battery, never fails open, and logs every mains cut.
- [flakehound](https://pctablet505.github.io/work/flakehound/): Static checks that find flaky pytest tests by their pattern, with rules for JAX, PyTorch and NumPy.
- [Daily Tracker](https://pctablet505.github.io/work/daily-tracker/): An offline Flutter habit tracker for Android with checksum-verified updates.
- [ATS optimizer](https://pctablet505.github.io/work/ats-optimizer/): A job-application pipeline that finds openings, builds each resume by selecting facts rather than inventing them, and fills forms with a browser robot.
- [ML-Guide](https://pctablet505.github.io/work/ml-guide/): An ordered route through AI and ML courses, in four stages.
- [Data structures in Python](https://pctablet505.github.io/work/data-structures/): A van Emde Boas tree with O(u) and O(n log log u) space versions, plus balanced trees, treaps, segment and Fenwick trees and skip lists.
- [A multi-agent engineering workflow on one workstation](https://pctablet505.github.io/work/agent-workflow/): A shared test scheduler, guard hooks and a context hub that let many coding agents work safely on one workstation.
- [Coursework projects](https://pctablet505.github.io/work/coursework/): Projects from MIT's machine learning course, CS50 AI and CS188, including Pac-Man search and reinforcement learning.

## Pages

- [Experience](https://pctablet505.github.io/experience/)
- [Open source](https://pctablet505.github.io/open-source/)
- [Writing](https://pctablet505.github.io/writing/)
- [Notes](https://pctablet505.github.io/notes/)
- [Demos](https://pctablet505.github.io/demos/)
- [Education and certificates](https://pctablet505.github.io/certificates/)
- [About](https://pctablet505.github.io/about/)

## Writing

- [A hot-expert cache made a 181 GiB model faster on one 96 GiB GPU](https://dev.to/pctablet505/a-hot-expert-cache-made-a-181-gib-model-faster-on-one-96-gib-gpu-4b03) (DEV, 2026-10-22)
- [Keras #22561 speedups: compiled Keras on PyTorch now generates text 11.6x faster](https://dev.to/pctablet505/keras-22561-speedups-compiled-keras-on-pytorch-now-generates-text-116x-faster-595m) (DEV, 2026-10-16)
- [Exporting Keras models to LiteRT](https://keras.io/guides/litert_export/) (keras.io, 2026-07-21)
- [Why a Keras model fails to convert to TFLite: 11 known cases, 5 traps](https://dev.to/pctablet505/why-a-keras-model-fails-to-convert-to-tflite-11-known-cases-5-traps-5bb0) (DEV, 2026-10-22)
- [Keras to TFLite on Android: wrong normalization and a backwards top-5](https://dev.to/pctablet505/keras-to-tflite-on-android-wrong-normalization-and-a-backwards-top-5-28m2) (DEV, 2026-10-22)
- [Two Android ML tests that could never fail, and a TFLite check that can](https://dev.to/pctablet505/two-android-ml-tests-that-could-never-fail-and-a-tflite-check-that-can-3bm8) (DEV, 2026-10-22)
- [Python projects are twice as likely as Java ones to leave most outside pull requests unanswered](https://porchlamp.hashnode.dev/python-projects-are-twice-as-likely-as-java-ones-to-leave-most-outside-pull-requests-unanswered) (Hashnode, 2026-10-23)

## Contact

- Email: work.rahul126@gmail.com
- GitHub: https://github.com/pctablet505
- LinkedIn: https://www.linkedin.com/in/pctablet505/
- DEV: https://dev.to/pctablet505
- Medium: https://medium.com/@pctablet505
- Hashnode: https://hashnode.com/@pctablet505
- Kaggle: https://www.kaggle.com/pctablet505
- X: https://x.com/pctablet505