Gopi Trinadh Maddikunta
Gopi Trinadh Maddikunta
From the lab
What I build and learn
Posts about the AI projects I build and the open-source work I do.
Interactive
Nova Bench: DGX Spark benchmarks
How fast can a $4,699 desktop run LLMs? I tested 11 models, from 8B up to 123B, on an NVIDIA DGX Spark and put the results in an interactive dashboard.
Emotion CNN: can a network read your face?
A CNN I trained on 35,000 faces to recognize 8 emotions. It runs live on your webcam, right in the browser. The post covers how I built it and how it works.
Deep dives
LLMs
Gemma 3 from scratch
I pre-trained a 164M-parameter Gemma 3 model from zero on the TinyStories dataset, so it writes simple children's stories. The post walks through the transformer internals I implemented, including Multi-Query Attention and RoPE.
See how I built it
Agents
Building BUD, my production AI agent
BUD (Build, Understand, Deploy) is a self-hosted AI assistant I built from scratch. It runs on 5 platforms at once with 24+ MCP tools, a 5-layer memory system, and real-time voice. No LangChain, no agent framework.
View the project
Open source
My GSoC 2025 journey
I was a Google Summer of Code 2025 contributor with the Scala Center, working on LLM4S, an open-source framework for tracing and monitoring LLM workflows. The post covers what I built and what the open-source work taught me.
Read the post