Exploring the frontiers of synthetic intelligence through curiosity-driven experimentation.
Eunice Labs is an independent AI research initiative run by Gagan, focused on understanding how large language models actually work — particularly their reasoning capabilities, agent reliability, and self-modifying systems.
Everything we build is open source. Documenting what works (and what breaks). The goal is simple: push the boundaries of what's possible with LLMs while making the research accessible to everyone.
We are currently obsessed with multi-agent systems and the reliability of tool calling.
Beyond the hype, we are investigating rigorous benchmarks for agentic behavior and building modular libraries for prompt optimization. The goal is to move from "it works sometimes" to "it works predictably."
Deep technical explorations of the architectures that power modern AI. Each post combines theory, mathematics, diagrams, and production-ready PyTorch code—breaking down complex concepts into buildable components.
From attention mechanisms to transformers
Before building agents and optimizing prompts, we need to understand the foundations. This series traces the evolution of neural sequence modeling—from basic seq2seq to the attention revolution that powers modern LLMs.
Each post includes mathematical explanations, architectural diagrams, and PyTorch implementations. Topics progress from encoder-decoder architectures through Bahdanau and Luong attention mechanisms, culminating in the complete Transformer architecture from "Attention is All You Need."
A comprehensive evaluation framework specifically designed to test the reliability, accuracy, and edge-case handling of LLM function calling capabilities across diverse domains.
Exploring novel architectures for long-term memory in autonomous agents. Investigating vector retrieval hierarchies vs. summary compression techniques for maintaining coherent persona state.
An open-source collection of highly optimized system prompts and few-shot examples for common reasoning tasks. Built to be modular, versioned, and easily integrated into Python codebases.
Important papers I have read and GitHub repos which contain code implementations — spanning foundational architectures, reasoning techniques, and model compression.
The foundational 2017 paper by Vaswani et al. that introduced the Transformer architecture.
Bahdanau et al. (2015) - The paper that introduced attention mechanisms to seq2seq models.
Luong et al. (2015) - Introduced global and local attention variants for improved efficiency.
Harvard NLP's line-by-line implementation guide with detailed explanations and code.
PyTorch implementations of all architectures covered in the Foundations series.