Final-year B.Tech (ECE) undergraduate  |  NIT Kurukshetra, India  |  Class of 2027
Collaborating with Sapienza University of Rome and the University of Tübingen

I build efficient, robust models for settings where compute, memory and privacy are real constraints, not afterthoughts. The common thread is models that understand the structure of the data instead of memorizing surface statistics.

• State space models & hyperspectral vision: Mamba/SSM architectures that respect the physics of the signal (SCAF, AbSpec-Mamba)
• 3D medical segmentation: sub-1M-parameter models for memory-constrained clinical hardware (LightMedSeg, RefineFormer3D)
• Event-based vision: treating event rate as a feature rather than a nuisance (CEGNN)
• Machine unlearning & privacy auditing: showing that "forgotten" classes still leak through a model's spatial and temporal output (ADVR, PACE)

"Efficient by design. Principled by necessity."

PyTorch 3D CNNs Spatial Anchoring BraTS ACDC

Learned spatial anchors guide the decoder without full-resolution feature maps at every layer. 84.8% Dice on BraTS, 93.5% on ACDC, 0.48M parameters, 14 GFLOPs. → arXiv

PyTorch Adversarial training Semantic segmentation Video Mamba attackers

Segmentation unlearning that relabels a class leaves a class-shaped hole in the output. ADVR removes the silhouette itself and is the only method leak-safe on both output and feature surfaces (ADE20K, Cityscapes). PACE extends this to video: averaging outputs over a ~2 s clip recovers the forgotten class from every published baseline, and the signal comes from cross-frame redundancy, not trajectory. → ADVR · PACE (arXiv) · PACE code

PyTorch Geometric GNNs Event cameras FiLM

The log event rate of each recording drives identity-initialised FiLM layers, trained with a two-stage rate curriculum: +14.0 points on DVS128 Gesture, +6.3 on N-Caltech101, and +34 points under synthetic speed shifts, for 49K extra parameters. Negative results documented. → arXiv · Code

Mamba Optimal transport Sinkhorn Hyperspectral

Differentiable optimal-transport grouping of spectral bands with per-group Mamba processors turns spectral redundancy into graceful degradation: +11.6% accuracy under 20 to 30% band dropout against Transformer baselines.

• AbSpec-Mamba / VolMamba v2: abundance-conditioned selective state transitions for physically grounded hyperspectral super-resolution (with Sapienza University of Rome)
• S6MOD: regret-driven adaptive routing for mixture-of-experts under distribution shift
• Selective Holographic State (SHS): sub-quadratic attention alternative based on low-rank holographic projections

• Reflara: replay-native execution-intelligence platform for autonomous AI agents. Deterministic record and replay of agent trajectories, written in Rust.
• Qten.ai: AI engineering intern, building agentic orchestration for grounded conversational content generation.
• SIGPLAN-M: operations team member, organizing international mentoring for programming-languages researchers.

• 📝 Research collaborations: efficient vision, state space models, event-based vision, machine unlearning and privacy auditing
• 🎓 Graduate research opportunities (2027 onward)
• 💬 Discussions: architecture design, SSMs, unlearning evaluation, CUDA optimization

📩 kavyansh.tyagi7@gmail.com

"Efficient by design. Principled by necessity."