Kest | Devpost

KEST

How we built it

We built Kest using Bun, TypeScript, OpenTUI, Nebius, NVIDIA Nemotron, and Tavily, with additional model provider support through Modal and Baseten.

Bun powers the application runtime, while OpenTUI provides an interactive terminal interface.

We integrated Nebius API access to NVIDIA Nemotron to power agent reasoning, research planning, and task execution.

We also integrated Tavily Search API to give our research agents access to information from the web, helping them discover relevant sources, conduct literature reviews, and gather evidence for their research objectives.

At the core of Kest is a multi-agent orchestration system that coordinates a director, research leads, workers, and independent reviewers. Each agent has specialized responsibilities while contributing to the same research objective.

We implemented autonomous agent loops, context management, research checkpoints, experiment tracking, persistent sessions, and budget allocation.

Kest supports up to three simultaneous model calls and three levels of delegation. Research artifacts are saved locally, allowing users to inspect findings, review experiments, and resume unfinished research.

By combining NVIDIA Nemotron for reasoning, Tavily for web research, and Bun for agent orchestration, we created a foundation for an autonomous AI research company.

GitHub: https://github.com/agkomyint/kest