Floki

Python · LangChain · MLflow · Langfuse

Inspiration

At DeepTek, debugging ML experiments meant diving into raw MLflow logs whenever anomalies appeared. Researchers spent hours parsing dense parameter tables and run metadata. Floki is a CLI agent that makes experiment auditing instantaneous: ask questions in natural language and get structured answers from local mlruns directories without leaving the terminal.

Demo

Sped up to focus on the core interaction loop and bypass standard MLflow Client initialization overhead.

Architecture and features

Agentic run discovery. Powered by LangChain, Floki routes user intents to targeted sub-tools. Natural language questions become structured queries that filter, match, and extract insights from local experiment directories.

Langfuse observability. End-to-end multi-turn session tracking and analytics monitoring under strict zero-cost infrastructure limits. Every routing decision, prompt iteration, and database payload is traced natively.

Terminal UI. A decoupled console UI pattern isolates backend inference from terminal rendering for a smooth, low-latency command-line experience.

Status

Active development. Current work expands native database access adapters, adds statistical analysis modules for multi-run delta comparisons, and refines intent-routing prompts for accuracy across thousands of localized experiment runs.