Designed Discrete Latent Operations: a 953M-parameter decoder-only transformer that replaces text chain-of-thought with a compact 6K+4-position latent schedule coupled to discrete state feedback from an isolated VM (>10x fewer tokens than text CoT). Reached 100% execution accuracy up to 195 sequential steps.

Architected Vadek OS, an AI operating layer for enterprises: hybrid retrieval (BM25 + 768-dim dense), automated source ingestion and taxonomy classification, role-based access control, DLP secret/PII redaction, and human-in-the-loop approval gates on AI-prepared actions.

Fine-tuned Llama 3.1 405B Instruct in bf16 (no quantization) on 8x NVIDIA H200 (141 GB each) using PyTorch FSDP with full sharding and LoRA (r=16, alpha=32); steady-state 44.8 s/step. Fixed a DTensor/bitsandbytes optimizer crash with fused AdamW and custom PEFT-based checkpoint loader.

Three-model remote-sensing suite built for the Smart India Hackathon: Siamese U-Net for bitemporal change detection (LEVIR-CD), dual-branch Optical-SAR cross-attention fusion (BigEarthNet), and a 7B remote-sensing VLM for VQA and asset-density estimation on 155 GB of Sentinel imagery.

Identified a failure class where Last-Write-Wins races in commercial LLM interfaces drop safety instructions mid-session; proposed deterministic guardrails (sequence enforcement, freshness validation, cryptographic verification). Published on Zenodo (doi:10.5281/zenodo.19350947).

Encrypted multi-client messaging system with authenticated, isolated sessions and zero-trust verification. High-throughput distributed telemetry with automated cipher streams and resilient peer verification. Published on Zenodo (doi:10.5281/zenodo.17875151).

Real-time license plate recognition and vehicle trajectory analytics for dense Indian road traffic using edge-optimized YOLOv11n and ByteTrack multi-object tracking.

Hackathon management platform that reached 500+ active users in under four months; owned architecture, go-to-market, and the Unstop partnership; backed by Google and Adobe for deployment and operations.

Selected as Google Student Ambassador; led campus initiatives and hands-on developer workshops on Google AI and cloud technologies, driving student engagement across machine learning and system engineering.