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. ๐ฆโ๏ธ
Detailed writeups on optimizing models, Quantization-Aware Training (QAT), and TensorRT execution are hosted directly on my repositories.
| Role | Company/Institution | Location | Dates |
|---|---|---|---|
| Data Science Intern | IIT Madras | Chennai, India | Dec 2024 - May 2025 |
| AI Development Intern | SkyGad | Remote | Aug 2024 - Dec 2024 |
- โก 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.
- ๐๏ธ 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. - ๐ค 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.
- ๐ NumPy Feedforward Neural Network (CS6910, Prof. Mitesh Khapra): Built a deep feedforward network from absolute scratch โ 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.
- ๐ฆ algofarm โ Rust Algorithm & Deep Learning Library: Systems-level ML library in Rust with PyO3 bindings, published to PyPI via
maturinand GitHub Actions trusted publishing. Includes hand-rolled data structures (open-addressingFarmHashMap, xorshift64* PRNG), sorting algorithms, and a reverse-mode autograd tensor engine โ no external crates. - ๐งฎ 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.
- ๐ CA RAG Backend: End-to-end retrieval-augmented generation backend for Chartered Accountancy exam content โ semantic chunking, Tesseract OCR, BGE-small embeddings, ChromaDB vector store, and a FastAPI layer with config-driven Ollama/Gemini LLM switching.
- ๐งฌ 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 โ positioned as workshop paper material (ENLSP @ NeurIPS).
- Google Developer Student Club (GDSC) IIT Madras โ Core Team Member & Technical Speaker ("Git and GitHub", "Dumping of Windows").
- Linux Community Lead โ IIT Madras.
- Al Horizons Conference โ Co-organized academic AI horizons forum alongside Prof. Sudarshan Iyengar.
- Smart India Hackathon โ Runner-up (2023).


