Open source. Open possibilities.

Reinforcement learning library for operational decisions—pricing, inventory, energy, queues, supply chains. 31 algorithms with typed interfaces, classical baselines, and reproducible benchmarks.

Antiserum is a local, open-source CLI scanner for text training data that flags possible planted poison such as trigger n-grams, label flips, duplicate injects, paraphrase overweight, hidden Unicode, instruction overrides, and known signature matches. It runs offline, writes deterministic receipts, supports CI/SARIF, and lets users confirm findings and contribute public signatures.

Variant-first CRISPR design framework that unifies SpCas9 nuclease, base editors and prime editors behind one typed interface, returning ranked candidate edits with calibrated uncertainty and population-aware off-target reports.

Defense-only fraud detection for cross-merchant identity reuse. Sentinel builds an identity link graph across merchants, scores payment events deterministically with explainable verdicts, and provides an analyst console, audit trail, and measured precision, recall, and false-positive cost in rupees.

AI Infra Bench evaluates frontier AI models on real-world AI infrastructure engineering tasks, starting with vLLM and covering CPU and GPU workloads including bugs, features, performance changes, refactors, and tests.

Reinfors is a high-throughput reinforcement-learning library that runs game dynamics, search, and episode orchestration in parallel Rust, while keeping Python-based networks and training entirely under the caller's control.

clinosim is a synthetic EHR data generator that creates clinically coherent HL7 FHIR R4 data through physiological simulation, supporting US and JP standards for AI training and testing.

FairMind is an open‑source AI governance platform providing bias detection, fairness testing, and compliance tracking for ML, LLM, and multimodal AI models. It offers evidence‑grade evaluation, model registry, and remediation tools.

An educational repository teaching supervised machine learning from first principles using Python, covering mathematical foundations and practical implementations.

Markovian is an experimental Haskell package for finite stochastic kernels, Markov reward/decision processes, and policies. It provides exact evaluation for bounded scopes including MRP/MDP evaluation, policy iteration, Bayesian inference, circuit representations, and finite open games. The package emphasizes mathematical rigor with separate modules for numerical, sampling, learning, and benchmarking capabilities.

Proving Ground is an independent, black-box benchmark and certification framework designed to objectively evaluate AI agents across twelve performance and safety dimensions.

PolyOCR Service is a multi-language OCR service built on PaddleOCR 3.x, supporting recognition for 78 languages, optional translation capabilities, and a standalone PaddleOCR-VL service.

Iris is a local-first MCP server that scores AI agent runs for quality, safety, and cost using 20 built-in deterministic rules, an optional LLM judge, and a web dashboard — all on your machine with no account or telemetry.

Shared native Rust compute core for NEAT-AI: the neat-core crate holds the scoring, loss and batched SIMD kernels used to run evolved neural networks, with optional rayon parallelism and a dual wasm32 and wasm64 WebAssembly bundle release pipeline.

Apollo is a Python package providing post-hoc statistical procedures with bias-aware corrections for common metrics in computational and ML/DL pipelines.

A tiny GPT language model (~540M parameters) designed to be trained from scratch on consumer laptop hardware, with pretraining on Fineweb-edu and chat finetuning capabilities.

An educational machine learning project that classifies online payment transactions as legitimate or fraudulent. It preprocesses a transaction dataset, encodes and scales features, then trains and evaluates Logistic Regression, Decision Tree, and Random Forest models with Python, pandas, NumPy, and scikit-learn.

GramGuidedURA provides code for semi-blind channel estimation in Unsourced Random Access (URA) using Gram-guided diffusion models.

Open LLM hallucination benchmark regarding the BNCC (Brazil's National Common Curricular Base). It measures how much models invent official codes and texts using an auditable methodology with raw data and CI.

comma-lab is a community and historical-record research workspace for the comma.ai video compression challenge. It provides the Task-Aware Compression library tac, lab operations tooling, exact replay custody, evidence grades, and OSS release hygiene.

NEAT-AI is a Deno/TypeScript neuroevolution library that evolves neural networks using WASM-accelerated activation, memetic evolution, error-guided structural discovery, and optional GPU-accelerated analysis via a Rust FFI extension.

Troy is a CLI tool for fine-tuning and preference-tuning LLMs locally on Apple Silicon Macs. Using MLX, it allows users to train models via a simple YAML config without cloud or CUDA requirements.

Lean 4 formalizations for response-based embeddings of black-box generative models, centered on the data kernel perspective space (DKPS) and multidimensional-scaling/spectral-perturbation infrastructure.

JouleWise is a benchmarking harness designed for the flexible power analysis of LLM inference workloads, focusing on auditable energy and latency measurement.

COSPA is a cost-performance benchmark for coding agents that evaluates models across coding, repository repair, and system administration tasks to determine value for money.

An open-source UFC fight prediction project: Elo ratings, a gradient-boosted ensemble and a multi-task neural network blended into one calibrated scorer, evaluated by expanding-window walk-forward over 2018-2026, with a Streamlit matchup explorer and a weekly self-updating data pipeline.

torch-tk is a toolkit designed to streamline the training, checkpoint management, and diagnostics of PyTorch models by adding structure to models and optimizers.