Skip to main content

​
What is Kapso?

Kapso combines iterative experimentation with a knowledge base of best practices, domain knowledge, and implementation patterns to discover and validate ML/AI code improvements. It automates the cycle of designing, running, evaluating, and refining solutions, eventually adapting the optimized solution for deployment on your chosen infrastructure.
Kapso Framework Architecture

​
Leeroopedia MCP Integration

Connect Kapso to Leeroopedia — a curated knowledge base of 1000+ ML/AI frameworks. Kapso agents use it during ideation and implementation to search docs, build plans, diagnose failures, and look up hyperparameter defaults. 
pip install leeroopedia-mcp
Sign up at leeroopedia.com for an API key ($20 free credit on sign-up). See the Leeroopedia MCP docs for setup details.

​
Core Workflow

from kapso.kapso import Kapso, Source, DeployStrategy

# Initialize Kapso with legal domain knowledge
kapso = Kapso(kg_index="data/indexes/legal_contracts.index")

# Research: Gather domain-specific alignment techniques
research_findings = kapso.research(
    "RLHF and DPO fine-tuning for legal contract analysis",
    depth="deep",
)

# Learn: Ingest knowledge from repositories
kapso.learn(
    sources=[
        Source.Repo("https://github.com/huggingface/trl"),
        *research_findings.repos(top_k=5),
    ],
    wiki_dir="data/wikis",
)

# Evolve: Build a solution through experimentation
# The agent writes evaluate.py to compute F1 score and signal when target is reached
solution = kapso.evolve(
    goal="Fine-tune Llama-3.1-8B on the CUAD dataset (https://huggingface.co/datasets/theatticusproject/cuad) for legal clause risk classification, target F1 > 0.85 on high-risk clause detection",
    output_path="./models/legal_risk_v1",
    context=[research_findings.ideas(top_k=10)],
)

# Deploy: Turn solution into running deployed_program
deployed_program = kapso.deploy(solution, strategy=DeployStrategy.MODAL)

# Cleanup
deployed_program.stop()

​
The Four Pillars

Evolve

.evolve() : Run iterative evaluation grounded experiments to build the optimum solution for a goal. Uses tree search, coding agents, and KG context to generate ideas and refine solutions.

Learn

.learn() : Ingest knowledge from repositories, past solutions, or written knowledge. Extracts patterns and best practices into the Knowledge Graph.

Research

.research() : Run deep web research to gather ideas, implementation references and releveant repositories. Returns structured findings you can pass to learning or use as context for evolving solutions.

Deploy

.deploy() : Turn a solution into running app. Supports local execution, Docker containers, or cloud platforms like Modal.

​
Core Components

ComponentDescription
KapsoMain API class — orchestrates research, learn, evolve, and deploy
OrchestratorAgentRuns the experimentation loop with budget tracking
Search StrategyExplores solutions via tree search (generic, benchmark_tree_search)
Coding AgentsPluggable code generators: Aider, Gemini, Claude Code, OpenHands
Feedback GeneratorValidates evaluation results and decides when to stop
Knowledge PipelineTwo-stage learning: Ingestors extract WikiPages → Merger integrates into KG
Knowledge SearchHybrid retrieval using Weaviate (semantic) + Neo4j (graph structure)
Repo MemoryEvidence-backed understanding of repository structure
DeploymentTurn solutions into running software: Local, Docker, Modal, BentoML

​
Supported Benchmarks

MLE-Bench

Solve Kaggle ML competitions with automated experimentation. Supports tabular, image, text, and audio problems.

ALE-Bench

Tackle AtCoder algorithmic optimization problems. Generates C++ solutions using optimization algorithms.

Quickstart

Get running in 5 minutes

Architecture

System design overview

Configuration

YAML config reference

​
Why Kapso?

  • Knowledge-Grounded: Solutions are informed by domain knowledge, not just LLM priors
  • Iterative Refinement: Tree search explores multiple approaches, learns from failures
  • Pluggable Components: Swap coding agents, evaluators, and search strategies
  • Full Lifecycle: From research to deployment in a single framework