Anup Jadhav

Prototype real-time voice agent for crisis support conversations. GPT-4o function calling drives its tools: grounding and breathing exercises, a safety assessment, and finding directions to the nearest hospital or police station with OpenStreetMap. Speech runs through Layercode (Deepgram STT, Rime TTS), with Twilio for phone calls and the backend on Cloudflare Workers.

A Model Context Protocol tool server that lets an AI assistant search code across GitHub, read files, and find repositories. Tools are described with JSON Schema so a model can discover and call them, and a browser client simulates an assistant using them.

Interactive tool to estimate GPU VRAM required to serve large language models on self-hosted hardware. Supports DeepSeek-V3.1, Llama, and Mixtral across A100, H100, B200 and other enterprise GPUs, with breakdowns of model weights, KV cache, and runtime overhead at multiple precisions.

Built an immersive Matrix-themed chatbot interface leveraging Heroku's AI inference and agent APIs. The project showcases modern AI capabilities while paying homage to the iconic film through its visual design and interaction patterns. Features include real-time message streaming, dynamic animations, and context-aware responses.

An EfficientNetB2 feature extractor computer vision model to classify images of food as pizza, steak or sushi. The model performs at 95%+ accuracy, and can classify an image at 0.03 seconds inference time per image.

Replicated the famous Vision Transformer paper using PyTorch to improve my understanding of Transformer architecture applied to vision problem(s).

A collection of articles and publications about AI Agents, software architecture, engineering management, and technology. Focused on making complex technical concepts accessible.