How bogoconic1/qgentic-ai works
Project type and purpose
• This is neither a web app nor a typical library but a research-and-development automation “pipeline” or “agentic R&D stack.”
• Purpose: Fully autonomous iteration on Kaggle competitions via collaborating LLM-driven agents (Starter, Researcher, ModelRecommender, Developer, Ensembler) under the control of an Orchestrator. It handles data ingestion, plan generation, code synthesis, execution, evaluation, and refinement.
File structure
• Top-level entrypoints:
• launch_agent.py (CLI)
• crawl_website.py / create_metadata.py (data prep)
• install.sh / node_install.sh (setup)
• project_config.py, config.yaml (configuration)
• agents/ : core agent logic (starter.py, researcher.py, model_recommender.py, developer.py, ensembler.py, orchestrator.py)
• prompts/, schemas/ : prompt templates and message schemas per agent
• tools/, guardrails/, utils/ : shared helper code (code execution, LLM tool calls, OOM retry, resource management, guardrail checks)
• task/ (runtime) : dynamic directory for competition bundles, intermediate outputs, logs, code artifacts
• tests/ : unit tests for each agent and tool
README for additional insights
• Provides a high-level “Architecture at a Glance” describing each agent’s responsibilities
• Lists key external integrations (Kaggle API, LLM endpoints, WandB, NVIDIA MIG, CPU affinity)
• Details on parallel baseline execution via ThreadPoolExecutor and resource pools
Detail without over-complication
• Every agent and its major interactions
• Concurrency model (parallel baselines)
• Two-stage DeveloperAgent flow (red-flags then SOTA) and how it loops
• Don’t fragment into too many micro-components; group helper modules under “Tools & Guardrails”