AI · Computer Vision · Physics-Informed ML
Building models that reason from first principles instead of memorizing patterns.
I'm an AI/Computer Vision researcher at IIIT Nagpur, working at the intersection of physics, cognition, and machine learning. My projects tend to share one thread: grounding models in the actual constraints of the physical world — sensor noise, light transport, quantum structure — rather than treating them as pattern-matchers over pixels.
- 🔬 Researching physics-anchored deepfake detection and mechanistic interpretability
- 🧬 Built an agentic AI tutoring platform with hierarchical memory and epistemic-state modeling
- 📷 Deep interest in sensor physics, noise modeling, and interpretable computer vision
- 🛠️ Bias toward building from scratch — no pretrained shortcuts when the goal is understanding
| Project | What it does | Highlight | Stack |
|---|---|---|---|
| 👻 Phantom Lens | Physics-anchored deepfake detection (PRISM V3) — validates whether video obeys real-world physics across 50 features / 4 pillars: PRNU sensor noise, light transport, compression forensics, rPPG cardiac signals. | AUC 0.97 in-distribution · 92% fake recall zero-shot cross-dataset | PyTorch · LightGBM · MediaPipe · SHAP |
| 🔬 Quantum-Structure SAE | Mechanistic interpretability: do classical transformers trained on quantum data (1D TFIM) encode quantum observables? Linear probing, TopK SAEs, activation patching, 5-part control battery. | Order parameter decodable at partial corr. 0.71 — but not load-bearing | PyTorch · PennyLane · scikit-learn |
| 🛡️ AEGIS | Agentic tutoring platform that thinks between sessions: 6 teaching agents, 4-layer memory stack, Ebbinghaus forgetting curve + SM-2, Theory-of-Mind modeling, always-on background cognition. | 6–10× memory compression · <5ms predictive pipeline | Next.js · TypeScript · Claude · D3.js |
| 🔭 PhysDenoiser | Image denoiser trained from scratch on physically realistic Poisson–Gaussian noise. Residual CNN, mixed L1+SSIM loss. No pretrained weights, no APIs. | 22.45 dB val PSNR · 30K-param CPU variant | PyTorch · NumPy · OpenCV |
Languages Python · TypeScript · C++ ML / CV PyTorch · OpenCV · scikit-learn · LightGBM · MediaPipe · NumPy · SciPy · PennyLane Interpretability Sparse Autoencoders · Linear Probing · Activation Patching · SHAP Web / Systems Next.js · React · Tailwind · Node.js · SQLite · Docker · Vercel
"Build models that reason from first principles — one physically-grounded system at a time."