The open-source specialty stack.

Towards fully automated open-ended scientific discovery [1.2.1]. An agentic framework that generates hypotheses, writes code, runs experiments, and drafts research papers autonomously.

Automating the Search for Artificial Life (ASAL) using foundation models. Replaces manual trial-and-error with agentic LLMs to discover novel lifelike cellular automata configurations.

Evolutionary optimization of model merging recipes. Automates the discovery of novel foundation model architectures by evolving weight-merging strategies across multiple open-source LLMs.

Open-ended and sample-efficient program evolution. Uses LLMs to iteratively mutate and evolve programmatic functions, demonstrating a novel approach to synthetic data generation.

An open-source agent harness designed to solve the ARC-AGI-3 reasoning challenge. Integrates a Python Read-Eval-Print Loop (REPL) enabling frontier models to interactively explore and solve abstract puzzles.

Learning through Autonomous Difficulty-Driven Example Recursion. Implements test-time reinforcement learning (TTRL) to let LLMs bootstrap their problem-solving skills without human supervision.

Joint Embedding Predictive Architectures for vision and video. Highly relevant for on-device property risk models that learn physics and structural dynamics without heavy supervised labels.

Comprehensive benchmark for assessing video generation and temporal consistency. Crucial for evaluating visual agents that reconstruct accident scenes or predict structural degradation over time.

A 4B-parameter unified multimodal model bridging image generation, editing, and reasoning. Ideal for agentic workflows where loss adjusters interactively query and modify damage reports on-device.

A lightweight knowledge-graph RAG framework. Bypasses multi-hop LLM costs, making it exceptionally cost-efficient for agentic retrieval across complex, interconnected commercial insurance policies.

The official PyTorch implementation of Google's Gemma models. Essential for self-hosting highly capable, open-weight reasoning engines that drive local underwriting co-pilots.

Dramatically accelerates LLM fine-tuning while slashing VRAM usage. The go-to toolkit for rapidly tuning bespoke actuarial reasoning models on a single consumer GPU.

Line-level text detection and OCR tailored for complex documents. Rapidly processes highly dense ACORD forms and schedules on-device without leaking sensitive PII to cloud APIs.

Converts PDF, EPUB, and MOBI to Markdown with exceptional accuracy. Provides the vital ingestion layer for RAG systems parsing sprawling corporate health plans and legacy wordings.

Open-source data pre-processing tools. The foundational pipe for transforming messy, raw broker submissions—from emails to PDFs—into clean chunks for agentic retrieval.

Blazingly fast DataFrames implemented in Rust. Replaces pandas for high-performance actuarial data manipulation, easily chewing through millions of exposure rows locally.

An in-process SQL OLAP database. Powers on-device data engineering, allowing agentic workflows to query massive claims datasets natively without a remote database connection.

An array framework for machine learning on Apple silicon. Unlocks phenomenal on-device performance for Mac-based field adjusters running offline diagnostic models.

Official inference framework for 1-bit LLMs. Reduces memory footprints drastically, making large-parameter reasoning accessible for low-power edge deployment in remote risk assessment scenarios.

Port of OpenAI's Whisper model in C/C++. The absolute gold standard for entirely offline, on-device transcription of distressed first notice of loss (FNOL) calls.

Inference of LLaMA models in pure C/C++. The engine powering the vast majority of decentralised, secure on-device LLM workflows handling sensitive insurance data.

Guarantees structural generation (JSON, Regex) from LLMs. Mandatory when an agent must parse a loss narrative and output a deterministic JSON schema directly into a rating engine.

Structured outputs using Pydantic. Streamlines the creation of robust validation layers for agentic extraction pipelines dealing with messy clinical and claims documentation.

A declarative framework that programs rather than prompts language models. Replaces fragile hard-coded broker negotiation prompts with self-optimising reasoning logic.

Transforms GUI screenshots into structured elements. Essential for Robotic Process Automation (RPA) agents that must navigate and scrape legacy claims management software autonomously.

A terminal-native, private medical assistant. Built for clinical workflows, this on-device agent tackles prior authorisation and coding audits without exposing health records.

State-of-the-art open coding models. Serves as a formidable on-device co-pilot capable of writing and executing custom actuarial python scripts directly on an analyst's workstation.

A barebones library for code-executing agents. Highly efficient for orchestrating on-device workflows where the LLM writes and executes Python to manipulate complex pricing datasets autonomously.

Framework enabling the development of applications using multiple conversing agents. Perfect for a setup where a 'Broker Agent' and an 'Underwriter Agent' negotiate terms automatically.

Houses the BGE-M3 model suite, handling multi-lingual, multi-granularity retrieval. Essential for global RAG agents searching through disparate regional compliance and regulatory frameworks.