TypeSafe AI
TypeSafe AI is an AI lab building machine-native, composable intelligence infrastructure for automation. Its mission is to make intelligence composable so it can be embedded as a dependable software primitive and catalyze a Cambrian explosion of intelligent software.
At a Glance
- Software developers and engineering teams
- Enterprise software and automation platforms
- AI agent builders
- Businesses needing reliable semantic judgment and decision automation
- +1 more
AI Tools by TypeSafe AI
(1)
Jev
Structured Decision AI Model
Discussions
No discussions yet
Be the first to start a discussion about TypeSafe AI
Latest News
TechCrunch: A new kind of AI model from a ChatGPT inventor is thrilling developers; coverage highlighted Jev's developer adoption, speed, cost and software-automation use cases.
LangChain published Building a Harness with Jev, documenting TypeSafe's model and its TypeSafeClassifier and agent middleware integrations.
TypeSafe announced the launch of System One Models and Jev, its first model for fast, structured decisions with calibrated confidence.
TypeSafe AI emerged from stealth with $40 million in seed funding led by DCVC and announced Jev in early access.
Products & Services
TypeSafe's flagship and first System One model. Jev evaluates typed questions against a state and returns structured decisions, probabilities and confidence rather than generated text. It supports Choice, Score and Noul question primitives and is available through the TypeSafe API in early access.
An API at POST /v1/systemone for sending text or structured state plus typed questions to Jev; independent questions are evaluated in parallel and return type-safe answers that application code can branch on.
Official Python library for the TypeSafe API, including TypeSafeClient and Choice, Score and Noul primitives.
Official TypeScript/JavaScript library for the TypeSafe API.
Market Position
TypeSafe positions System One/Jev as a complement to, rather than a general replacement for, generative LLMs: Jev handles bounded, structured decisions, routing, scoring and verification, while generative models handle open-ended reasoning and prose. Its differentiation is calibrated probabilities, type-safe outputs, parallel sampling, and claimed substantially lower latency and cost than frontier LLMs. Likely competitive alternatives include general LLM APIs from OpenAI, Anthropic and Google for classification and structured-output tasks, and agent/model-routing frameworks that use those models; TechCrunch reported that competitors are likely to emerge around this decision-model category.
Leadership
Founders
Diogo Almeida
Co-founder and CEO; former OpenAI researcher who co-invented reinforcement learning from human feedback (RLHF) and worked on InstructGPT, ChatGPT and GPT-4; previously at Google Brain.
Sasha Sheng
Co-founder and COO; former research engineer at Meta/FAIR, where she worked on News Feed, AI Experiences and AI Research; published at NeurIPS and ECCV.
Erik Gafni
Co-founder and CTO; repeat founder of Ravel, a multimodal AI company for DNA sequencing; early employee at Invitae and Freenome; inventor with publications and patents specializing in production AI systems.
Executive Team
Diogo Almeida
Co-founder and CEO
Former OpenAI researcher and co-inventor of RLHF and InstructGPT; previously at Google Brain.
Sasha Sheng
Co-founder and COO
Former Meta/FAIR research engineer working on News Feed, AI Experiences and AI Research; NeurIPS and ECCV author.
Founding Story
Diogo Almeida says he started TypeSafe after concluding that models optimized for human-preference chat still required humans in the loop and were difficult to build into reliable software. After leaving OpenAI in 2024, he and co-founders Erik Gafni and Sasha Sheng spent two years in stealth pursuing machine-native AI: intelligence that can run quietly in software, return calibrated decisions, and be composed like a dependable software primitive.
Business Model
Revenue Model
Usage-based API access priced by input tokens; output tokens are free. Higher rate limits are available on custom and enterprise plans, and TypeSafe also offers enterprise access.
Pricing Tiers
Documented rate limits are 250,000 tokens per second and 1,200 requests per minute; higher limits are available on custom and enterprise plans.
Target Markets
- Software developers and engineering teams
- Enterprise software and automation platforms
- AI agent builders
- Businesses needing reliable semantic judgment and decision automation
- Industries such as insurance underwriting and other workflows where calibrated confidence and human escalation are important
- AI-powered workflows and smart if-statements
- Classification, routing, scoring and extraction inside application code
- Model routing for selecting an appropriate generative model
- Real-time applications where latency matters
- Map-reducing over large datasets to produce features and insights
- Verification, guardrails and jailbreak detection for LLM prompts, traces and outputs
- Vercel
- LangChain
- Bryo AI