TypeSafe has officially emerged from stealth with Jev, an AI model designed to handle structured probabilistic decisions directly within software systems. The company was founded by OpenAI veteran and ChatGPT co-inventor Diogo Almeida after two years of development.
Unlike conventional large language models that generate responses token by token, Jev is designed to skip text generation altogether. The model accepts unstructured state as input and returns type-safe structured values through a single parallel query.
The approach is intended for software environments where predictable outputs, strict schemas and fast execution are more important than natural-language interaction.
Jev Uses Parallel Sampling Instead of Token Generation
TypeSafe developed Jev around a training approach called Reinforcement Learning for Calibrated Decisions (RLCD). Rather than optimizing primarily for conversational responses or individual verifiable tasks, the system is trained to produce probability estimates that correspond with the accuracy of its decisions.
The architecture also replaces conventional autoregressive generation with a hardware-aware parallel sampling system. Multiple structured values can be evaluated simultaneously, allowing Jev to make decisions without sequentially generating text.
Outputs are restricted to predefined schemas, which can reduce the need for parsing layers and additional guardrails commonly used when integrating traditional language models into production workflows. The sampler can evaluate selections involving up to 255 discrete options through a two-stage scoring and selection process.
According to TypeSafe’s internal testing, Jev has recorded end-to-end response times ranging from 70 to 500 milliseconds. The company also reports workflow evaluations achieving execution speeds as much as 193.6 times faster than selected conversational model baselines.
Lower Inference Costs Target High-Volume AI Workloads
TypeSafe is positioning Jev for applications where AI decisions need to run continuously at scale. The company says input processing costs $0.042 per million tokens, while output tokens are not metered because the system generates structured states rather than conventional autoregressive text.
This architecture could make the model suitable for workloads involving large numbers of small, automated decisions, particularly where traditional LLM inference introduces unnecessary latency or processing overhead.
Almeida describes the system as a function-like interface: unstructured state goes in, while typed probabilistic decisions come out.
Jev Tested on Games, Web Traversal and Large-Scale Data
TypeSafe has demonstrated Jev across several decision-heavy environments. One test involved a real-time Doom bot processing game-state branches at 10 queries per second, with reported operating costs of roughly $7 per hour.
Another experiment used Wikiracing to test navigation through dense encyclopedia link structures. Jev selected links across the traversal tree while avoiding the generation of unnecessary intermediate text or hallucinated paths.
The company says its broader target applications include real-time feature extraction, large-scale data processing, output verification and automated branching logic.
TypeSafe has now opened early developer access and begun onboarding engineering teams from its existing deployment waitlist.