Projects.

Deep learning model that detects glaucoma from retinal scans with high accuracy.

What I've done

• •Built a clinical-grade AI system that detects glaucoma from eye images with 94.5% AUC and 87.1% accuracy, beating traditional ML models by 15%
• •Used ResNet-18 for image classification and U-Net for optic disc/cup segmentation, plus Grad-CAM to show doctors why the AI made its decision.
• •Used ONYX instead of Pytorch for better performance.

A high-performance RAG platform for querying YouTube channel content with semantic precision and keyword-exact retrieval.

What I've done

• •Implemented a Hybrid Retrieval system using Reciprocal Rank Fusion (RRF) to combine Vector (Semantic) and BM25 (Keyword) search, significantly improving retrieval accuracy over standard RAG.
• •Engineered a lightweight, CPU-optimized embedding pipeline using FastEmbed (ONNX), reducing memory overhead by 80% compared to traditional PyTorch-based transformers.
• •Built an automated ingestion pipeline with a background scheduler and SQLite metadata tracking to handle multi-video processing and transcript extraction (yt-dlp) asynchronously.

A high-performance local AI model benchmarking platform for profiling latency, throughput, cold-starts, and parameter isolation.

What I've done

• •Designed a local LLM benchmarking pipeline utilizing Ollama to profile TTFT, throughput (TPS), and latency across Llama 3.2, Qwen 2.5, and Phi 3 Mini.
• •Engineered comparative analytics dashboards to isolate temperature behaviors (0.0 vs 0.7) and analyze GPU cache-warming and WSL2 cold-start performance.
• •Built a FastAPI backend with SQLite relational storage and a glassmorphic React frontend featuring side-by-side Recharts visualizations.