computer vision, edge AI engineer + ML developer with a passion for high-frequency charts & neo-retro bikes

I optimize real-time deep learning pipelines and deploy computer vision systems on low-power edge hardware. The objective is to design hyper-lightweight, high-throughput model architectures running at maximum physical frame rates. 🦀⚙️

Detailed writeups on optimizing models, Quantization-Aware Training (QAT), and TensorRT execution are hosted directly on my repositories.

• ⚡ Real-Time Traffic & Surveillance Pipeline (IIT Madras): Deployed directional-aware YOLOv5/v8 object detection on NVIDIA Jetson edge devices. Optimized inference throughput via TensorRT layer fusion and precision calibration. Integrated FAISS-based access tracking and automated reporting.
• 👁️ Face Recognition & Temporal Tracking System: Real-time facial validation pipeline linking YOLOv8, ByteTrack, and ArcFace (InsightFace buffalo_l) embeddings queried against a dynamic FAISS index (IndexFlatIP, L2-normalized). Stabilized identity across video sequences via K-means-clustered enrollment centroids and sliding-window temporal label smoothing.
• 🤖 RAVA (Retrieval-Augmented Virtual Assistant): Memory-enabled conversational agent built on LangGraph and the Google Gemini API, with a vector index powering per-user, localized long-term context retrieval.
• 📐 NumPy Feedforward Neural Network (CS6910, Prof. Mitesh Khapra): Built a deep feedforward network from absolute scratch — zero PyTorch/TF autodiff. Hand-implemented all 6 optimizers (SGD, Momentum, NAG, RMSProp, Adam, Nadam) and ran WandB sweeps to benchmark accuracy on Fashion-MNIST.
• 🦀 algofarm — Rust Algorithm & Deep Learning Library: Systems-level ML library in Rust with PyO3 bindings, published to PyPI via maturin and GitHub Actions trusted publishing. Includes hand-rolled data structures (open-addressing FarmHashMap, xorshift64* PRNG), sorting algorithms, and a reverse-mode autograd tensor engine — no external crates.
• 🧮 SIMD Matrix Multiplication Engine (Rust): Custom GEMM kernel using NEON SIMD intrinsics, exposed to Python via PyO3/maturin and benchmarked against PyTorch CPU on Apple Silicon. Diagnosed the residual performance gap to Apple's AMX coprocessor being inaccessible from userspace NEON code.
• 📄 CA RAG Backend: End-to-end retrieval-augmented generation backend for Chartered Accountancy exam content — semantic chunking, Tesseract OCR, BGE-small embeddings, ChromaDB vector store, and a FastAPI layer with config-driven Ollama/Gemini LLM switching.
• 🧬 Knowledge Distillation with Matryoshka Representations: Distilled a 6-layer, 22.4M-parameter encoder-only student model from BAAI/bge-large-en-v1.5, combining dynamic piecewise-linear (DY-ReLU) activations with Matryoshka representation learning — positioned as workshop paper material (ENLSP @ NeurIPS).

• Google Developer Student Club (GDSC) IIT Madras — Core Team Member & Technical Speaker ("Git and GitHub", "Dumping of Windows").
• Linux Community Lead — IIT Madras.
• Al Horizons Conference — Co-organized academic AI horizons forum alongside Prof. Sudarshan Iyengar.
• Smart India Hackathon — Runner-up (2023).