Projects

A mini-Triton JIT compiler for LLM GPU kernels: maps Python block-level semantics to GPU thread layouts via NVRTC. Reaches bandwidth parity with OpenAI Triton on memory-bound kernels (Softmax, LayerNorm) and 110 TFLOP/s on FP16 tensor-core matmuls, using custom PTX intrinsics, swizzled shared memory, and software pipelining.

An adaptive neural-architecture-search pipeline built for out-of-the-box generalization on novel datasets under a strict 10-minute compute budget. Dynamic, budget-aware capacity scaling guarantees convergence without OOM faults: zero crashes, and it outperformed every baseline benchmark in Phase 2.

A multi-agent survival sandbox with deterministic causation-gating, built to separate structural hoarding from genuinely emergent deception in LLMs. Stripping prompt disposition collapses reported deception by 85%, exposing severe measurement confounds in current agentic evaluations.

SAIR Foundation challenge (Terence Tao, Damek Davis): distilled equational-implication reasoning over magmas into a 7.44 KB two-phase decision-procedure cheatsheet (60.2% accuracy at $0.00037/problem), smaller and cheaper than the 1st-place entry. Read the writeup.

Competing in François Chollet's ARC-AGI-3 - one of AI's hardest benchmarks for fluid intelligence and abstract reasoning. Exploring learning-based and program synthesis approaches beyond random baselines.

An interactive gallery exploring quantization techniques for large language models - comparing methods, quality metrics, and model sizes.

Lightweight reinforced reasoning model for Text-to-SQL using GRPO RL and DeepSpeed distributed training on a 7B backbone. SOTA on the BIRD benchmark, outperforming 236B+ models.

Real-time ASL hand sign detection using OpenCV, cvzone HandDetector, and a Keras classification model trained on 7 signs.

AI assistant for object identification using computer vision and NLP. Identifies tourist attractions, animals, plants and converts descriptions to audio via TTS.

Real-time privacy protection system using facial recognition to blur faces in video streams. Built with OpenCV and face_recognition library.

AI-based attendance tracking system using face recognition via mobile camera. Supports multiple faces, automated email notifications, and real-time tracking.

Web application using ML algorithms to analyze crime patterns and predict criminal activity from law enforcement and social media data.

Standalone forecasting application with multiple ML models, data upload, forecast generation, and exportable outputs. Trained on universal demand data.

Web app that identifies dishes from photos and suggests recipes. Built with Streamlit, Keras (InceptionV3), and Selenium for automated food delivery search.

Interactive simulator for FCFS, Priority, and Round Robin scheduling algorithms. Built with Streamlit, Pandas, and Matplotlib for visualization.