ML engineer and open-source contributor to Keras — I enjoy making machines a little smarter every day, and making the smart ones run fast on small devices. Previously at Qualcomm and Samsung R&D, building face recognition, computer vision, and sensor-driven systems for millions of devices. Based in India 🇮🇳.
🌐 pctablet505.github.io · 💼 LinkedIn · 📈 1,000+ contributions in the last year
100+ merged pull requests in 70+ open-source projects, 60+ of them in the Keras ecosystem (keras · keras-hub). GitHub keeps the live count: all merged pull requests →
Merged work I'm proud of:
- keras#22758 — LiteRT export for the PyTorch backend: on-device deployment of PyTorch-trained Keras 3 models
- keras-hub#2132 — Llama 3.1 implemented in KerasHub
- keras#22362 — serialization fix: custom models with sublayers in nested lists failed to save/load
- keras#22998 — torch-backend
SymInthandling inconvert_to_tensorandslice
In review right now: 40+ pull requests, including vLLM, Keras and keras.io — open pull requests →
flowchart LR
A["Frontier model<br/>(Gemma, …)"] --> B["Keras 3<br/>multi-backend"]
B --> C["LiteRT / LiteRT-LM<br/>quantize + export"]
C --> D["📱 On-device<br/>fast · private · offline"]
Edge & on-device AI · model optimization (quantization, latency, memory) · multi-backend Keras internals · computer vision & applied ML that actually ships.
Outside of Keras, my main project is RL Alpha Labs — a research-stage, JAX-native lab for algorithmic trading on Indian equities: PPO/DQN agents tested across 746 NSE instruments on 3.5+ years of market data, 128 vectorized environments for parallel rollout, and full portfolio simulation with risk controls; no live capital is deployed. The code is private; the research, field notes and benchmarks are at rlalphalabs.com →
🕶 The part you can't see here: my earlier work was face recognition and face-related solutions — plus computer-vision and camera/sensor problems — at Samsung R&D and Qualcomm. Real-time face auth serving 10,000+ registered faces on-device, embedding search cut from 1,200 ms → 87 ms, sensor-based drop detection at 95% less power. All closed-source, so GitHub only shows my Keras side.
| Project | What it is |
|---|---|
| Porchlamp 🏮 | Find the projects that leave the light on — 11,000 open-source projects ranked by what happens to a stranger's pull request, not by stars · latest report · live ranking · repo |
| RL Alpha Labs 🔒 | Research-stage, JAX-native RL trading research on Indian equities — PPO/DQN over 746 NSE instruments, 128 vectorized envs, portfolio simulation with risk controls · private code, research and notes |
| Servedeck | A systemd control plane, an OpenAI-compatible /v1 gateway and a live dashboard for native vLLM on one Linux GPU workstation · docs |
| GLM-5.3 on one GPU | GLM-5.3-Flash (320B total, 18B active, NVFP4) on one RTX PRO 6000 Blackwell — the best result we know of on a single RTX PRO 6000, with the full method |
| flakehound | Hunt flaky tests before they bite: static flaky-pattern detection and run-history scoring for pytest, ML-testing aware (JAX, PyTorch, NumPy) |
| ats-optimizer 🔒 | Truthful, human-in-the-loop job-application automation — feed-based discovery, knowledge vault, tailored resumes, ATS analysis · private repo |
| litert-demo | End-to-end export of KerasHub Gemma3 → LiteRT .tflite / LiteRT-LM .litertlm |
| gemma-tflite-android-demo | Gemma running fully on-device on Android |
| jax-windows-cuda-build | CUDA build scripts, patches & pre-built JAX wheels for Windows |
| daily_tracker | Cross-platform daily task tracker with Google Drive sync & auto-updates |
| ML-Guide | Curated AI/ML learning path — courses, books, study notes |
- Better performance from smaller models — quantization, distillation, smarter kernels
- Synthetic data generation techniques
- Building a gallery search system