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 →
pctablet505Rahul Kumar, machine learning engineer, India
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.
10.8 → 15.0 tokens/sdecoding speed, from plain offloading to the shipped setup● checkedGLM-5.3-Flash on one GPU →70% → 90%+damaged barcodes read correctly on the damaged synthetic test set, a commercial scanning SDK against this reader○ reportedReading damaged barcodes on phones →millions → under 22model parameters○ reportedPhone drop detection in under 22 parameters →69 of 70capital-and-breadth tiers in which an index fund beat holding the stocks directly, after costs and tax● checkedRL Alpha Labs →
What I work on
One habit in every area: measure what each change bought
Performance and inference
Big models made to fit and run fast on phones, NPUs and one GPU, and the frameworks that ship them.
Computer vision
Seeing on the device: damaged barcodes, faces, attention, documents.
Sensor and behaviour ML
Reading motion and touch: falls, crashes, and who is holding the phone.
Projects and tools
An independent research lab, and the tools built along the way.
Selected work
The work I would show first
Each card opens the full story: the problem, what I did, the result, and where the number comes from.
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 →
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 →
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 →
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 →
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 →
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 →
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 →
Demos lab
Try the ideas in your browser
Rebuilt from scratch with public techniques. Each page shows how it works today; live versions are on the way.
Damaged barcode reader
Point a camera at a scuffed barcode, or upload a photo, and watch it get found, cleaned up and read, or refused when the damage is too much.
Which hand are you using?
Tap and swipe a few times on your phone and the page guesses which hand holds it: the first signal in telling users apart by behaviour.
Face and attention on camera
Live face landmarks with attention and expression signals from your camera.
Open source
Merged upstream, in the open
Most of the work is in Keras and KerasHub; the rest is fixes across the Python and ML tooling people use every day.
Writing
Articles
- Python projects are twice as likely as Java ones to leave most outside pull requests unanswered ↗23 Oct 2026
Hashnode
- A hot-expert cache made a 181 GiB model faster on one 96 GiB GPU ↗22 Oct 2026
DEV
- Why a Keras model fails to convert to TFLite: 11 known cases, 5 traps ↗22 Oct 2026
DEV
- Keras to TFLite on Android: wrong normalization and a backwards top-5 ↗22 Oct 2026
DEV
Notes
The knowledge base
- The eight bugs between a fresh checkout and correct output ↗
GitHub
- Backtest traps that fooled us ↗
rlalphalabs.com
- Optimizing GLM-5.3-Flash on one GPU ↗
GitHub
- What trading really costs on NSE ↗
rlalphalabs.com
Contact
Working on on-device ML, computer vision or fast inference?
I am glad to talk about roles, collaborations and hard problems where a model has to fit.