Implementing arXiv 2504.19874 (TurboQuant) on Qwen 32B to compress the KV cache via random rotation + scalar quantization, targeting 4x memory reduction for long-context inference on a single H100. Built custom Triton kernels for fused rotate-quantize-dequantize on the attention path; benchmarking against vanilla FP16 KV cache on perplexity and throughput.

Built a virtual world where blank GRU-based agents evolve from scratch via genetic algorithms with no reward shaping, reaching 86 generations of emergent foraging and survival behavior. Designed a phased curriculum (perception → action → memory → social) to study how cognition emerges in minimal artificial-life systems.

Building an OLMo language-model wrapper for DeepChem, bringing fully open-source LLMs (weights + training data + code) to the cheminformatics ecosystem for molecule property prediction and chemical reasoning. Contributing to a 2.5k+ star scientific ML library used by pharma research labs; PR live and under maintainer review.

Developed a stealth desktop app for AI-powered coding/interview assistance with advanced process hiding. Integrated Google Gemini and OCR for real-time speech/screen analysis and instant responses. Designed modular, cross-platform (Win/macOS/Linux) architecture and intelligent caching.

Implemented Group Sequence Policy Optimization (GSPO) algorithm from Qwen Team research, achieving superior stability with 50-75% clipping rates vs 0.01-0.02% for baseline PPO/GRPO methods on reasoning tasks. Engineered complete knowledge distillation pipeline from DeepSeek-R1 to Qwen-1.5B architecture, incorporating 8-bit optimization and gradient checkpointing for memory-efficient training on H100/RTX hardware.

Developed production AI system for Andhra Pradesh Police processing 1000+ emergency calls daily, implementing speech-to-text, sentiment analysis, and priority classification. Built real-time geospatial analysis and emergency dispatch optimization, reducing average response time by 25% through intelligent resource allocation.

Developed multimodal AI system transforming 2D engineering diagrams into 3D CAD models with 92% geometric accuracy using computer vision and LLM integration. Built robust feature extraction pipeline handling technical drawings, sketches, and annotations with advanced OCR and geometric reasoning capabilities.

Reproduced DeepSeek R1 Zero achieving 90% functional parity, implementing advanced RL techniques for LLM self-verification on complex reasoning tasks. Engineered distributed training pipeline using VLLM backend and Flash Attention 2, enabling cost-effective training of 3B parameter models on consumer hardware with 40% cost reduction.

Engineered production-ready pose transfer system integrating 5 ControlNet models, achieving 90% pose accuracy on COCO-Pose benchmark with real-time inference. Implemented advanced prompt engineering pipeline with dynamic negative prompting and attention guidance, improving generation quality by 40% and reducing artifacts by 65%.

Built intelligent CLI agent leveraging Google Gemini LLM for natural language to system command translation, enabling conversational OS interaction with 95% command accuracy. Engineered autonomous security auditing system with real-time monitoring, policy violation detection, and automated threat response capabilities.