Third-year B.Tech (AI) student at Amrita Vishwa Vidyapeetham, Coimbatore. I work on understanding and editing what language models know, on image deblurring, and on local-first AI that runs on one laptop GPU.
- Interpretability: cross-lingual causal tracing and rank-1 editing in Llama-3.2-3B and Gemma-2-2B (babel-effect, paper draft in preparation).
- Computer vision: WaveFusion-Net, a dual-branch spatial-wavelet deblurring network (9.48M parameters), IEEE NIGERCON 2026 (code).
- Shipped product: Coast, a peer-to-peer ride-matching Android app built with React Native, Expo and Supabase (coastapp.in).
| Project | Change | Status |
|---|---|---|
| microsoft/PyRIT | Add a request timeout to remote dataset downloads | Merged |
| NVIDIA/NeMo-Agent-Toolkit | Stop logging the expanded config when --override is used, which could expose secrets |
Merged |
| mozilla-ai/otari | guardrails validate warns when a when_changed glob matches no file |
Merged |
| TransformerLensOrg/TransformerLens | Reject negative temperature and non-positive repetition_penalty in sample_logits |
Merged |
| microsoft/PyRIT | Match % and _ literally in the objective and scenario name filters |
Open |
| EleutherAI/delphi | Keep n_train examples when train_type is mix (float flooring dropped examples) |
Open |
| SylphAI-Inc/AdalFlow | Make append_to_jsonl safe across threads |
Open |
| arc53/DocsGPT | Read markdown, rst and csv files as UTF-8 | Open |
| Project | What it is | Result |
|---|---|---|
| babel-effect | Do LLMs store facts in language-specific circuits? Pre-registered hypotheses, 1059 claims in a verifiable manifest | English-Spanish causal-profile alignment above the shuffled null (z = 4.0 Llama, 4.8 Gemma) |
| WaveFusion-Net | Spatial-wavelet fusion for blind image deblurring | 9.48M parameters, SHA256 split-leakage checks |
| Knee OA grading | ResNet ensemble trained from scratch, with Grad-CAM | 65.7% accuracy, kappa 0.813 |
| Multi-Scale Attention U-Net | Image deblurring with CBAM attention and a Charbonnier + FFT loss | 28.1 dB PSNR, 0.815 SSIM on GoPro |
| HospiCall-AI | Hospital voice receptionist: Whisper, Llama 3.1 and TTS, fully local | Demo flow in the README |
| liftgain-tracker | Workout and nutrition tracking on PostgreSQL: 3NF schema, PR-detection trigger, window-function views | FastAPI backend, React Native app |
| signum-job-intelligence | Job-market skill tracker on TimescaleDB with a custom Redis bitmap index | Real data sources, no synthetic data |
Python, PyTorch, TransformerLens, FastAPI, Ollama, React Native, Expo, TypeScript, PostgreSQL, Supabase.