Approximate Labs
Approximate Labs is an independent research lab studying the efficiency gap between human and machine reasoning. It builds benchmarks, tools, and architectures intended to close that gap, with a particular focus on data, agents, and model reasoning.
At a Glance
- AI and machine-learning developers
- Model-serving and inference infrastructure teams
- Agent and multi-agent application builders
- Reinforcement-learning and process-supervision researchers
- +2 more
AI Tools by Approximate Labs
(1)Affordance
Open Source TSX Browser for LLMs
Discussions
No discussions yet
Be the first to start a discussion about Approximate Labs
Latest News
Affordance, a browser for models, is open source
Shikumi: persistent GPU processes communicating through typed channels
A few Craftax solves
Pencil Puzzle Bench: a benchmark for multi-step verifiable reasoning
Products & Services
Open-source Rust engine and browser for models. Developers write agent applications in a documented TypeScript/JSX subset; Affordance renders the app into the exact token IDs a model reads, exposes tools, manages state/history, and runs handlers. Version 0.1 is a beta under Apache-2.0.
Benchmark and Gym-compatible environment for multi-step reasoning on pencil puzzles. It uses deterministic, step-level verification and supports direct and agentic evaluation with verifier feedback.
GPU runtime architecture built around persistent processes that run at their own rate and communicate through typed channels; demonstrated in a single-kernel training, inference, and game workload.
Open-source dataset of 627 million tables with context, created for training large/data-focused models.
Market Position
Approximate Labs positions itself as an independent research lab focused on reasoning efficiency rather than as a conventional AI SaaS vendor. Affordance differentiates from prompt-component systems such as Priompt and Microsoft's prompt-tsx by rendering through to model token IDs, incorporating tools and the agent loop, and supporting both agents and model-server integration. In data-focused AI, Caplight identifies PandasAI, Mito, and Julius AI as comparable companies.
Leadership
Founders
Justin Waugh
Founder/CEO of Approximate Labs; previously Chief Scientist and Chief Architect at Unsupervised, and a data scientist at Divirod. He holds a PhD in Physics from the University of Colorado Boulder and a BS in Electronics Engineering from Virginia Tech.
Executive Team
Justin Waugh
Founder and Chief Executive Officer
Previously Chief Scientist and Chief Architect at Unsupervised, data scientist at Divirod, and a postdoctoral researcher/research assistant at the University of Colorado Boulder; PhD in Physics.
Founding Story
Approximate Labs was established in 2022 by Justin Waugh to build AI systems that work natively with data and to investigate the efficiency gap between human and machine reasoning. Its earlier work focused on tabular-data systems, including the TabLib dataset, Sketch, Julyp, and TableGen; its current work has broadened to verifiable reasoning benchmarks, model-agent interfaces, and GPU/runtime architectures.
Business Model
Revenue Model
Approximate Labs is described by Caplight as a B2B business with subscription-SaaS and freemium characteristics. Its current Affordance offering is open source under Apache-2.0, while the lab's earlier Julyp product was a data-focused AI assistant.
Target Markets
- AI and machine-learning developers
- Model-serving and inference infrastructure teams
- Agent and multi-agent application builders
- Reinforcement-learning and process-supervision researchers
- Data-science and business-intelligence users
- Organizations working with tabular data and natural-language data interfaces
- Building custom coding agents, research agents, memory systems, and multi-agent fleets
- Training and reinforcement learning where the exact tokens seen and generated by the model must be preserved
- Serving and integrating model applications with SGLang, vLLM, and chat endpoints
- Long-context prompt layout, cache-aware history management, and agent harness experimentation
- Process-supervision, verifier-based training, tool use, and error recovery research
- Tabular data analysis, natural-language data interfaces, dashboards, and data-focused assistants
- Multiplet, whose agent fleets are described as running on Affordance