← Back to feed

TypeSafe AI launched Tuesday with $40 million in seed funding led by DCVC. The company’s first model, Jev, is now in early access.

TypeSafe was founded by Diogo Almeida, who spent time at OpenAI developing the methods behind ChatGPT’s instruction-following capabilities — specifically, the reinforcement learning from human feedback work that made language models useful at dialogue. He left OpenAI, he says, after concluding that chat models were not the right path to meaningful automation.

The company has been in stealth for two years.

What System One Models Are

TypeSafe’s framing: current LLMs are optimised for human preference on sequential chat messages. They output strings, which are flexible but require parsing, validation, and carry the risk of going off-script.

System One Models, TypeSafe’s term for its new class, output typed structured values. Possible outputs and their structure are defined in advance. The model cannot make type errors. It cannot generate hallucinated strings because it does not generate strings at all.

Each output comes with calibrated probabilities and confidence scores. TypeSafe calls this RLCD — Reinforcement Learning for Calibrated Decisions — the training method that produces epistemically honest probability estimates rather than optimising for human preference ratings.

The intended use case: software-native automation. Jev is designed to sit inside pipelines and make structured decisions based on program state, not hold conversations.

Almeida describes the intended interface: “unstructured state in, typed probabilistic decisions out.”

Jev Performance Claims

TypeSafe claims Jev achieves “similar levels of intelligence on System One tasks” compared to existing LLMs, while being two orders of magnitude faster and more efficient. The company has not published independent benchmark results alongside the launch.

The two-orders-of-magnitude efficiency claim is specific to System One tasks — structured decision-making with typed outputs — not general reasoning or chat. TypeSafe’s case is that it is solving a different problem than frontier labs, not outperforming them on shared benchmarks.

The Funding and Investor Context

DCVC (Data Collective Venture Capital) led the $40 million seed round. DCVC invests primarily in deep tech with a scientific or engineering thesis. The seed size is large relative to typical pre-product rounds, reflecting two years of prior development and a technical team with frontier lab pedigree.

The Market Thesis

The company’s argument is that the automation bottleneck is not intelligence — it is interface. LLMs produce text that downstream software has to interpret, validate, and handle failure cases for. TypeSafe builds models whose outputs are already typed, validated, and probability-weighted, removing the parsing layer entirely.

Whether that thesis proves out depends on how broadly “automation” can be decomposed into typed decisions. Complex multi-step workflows often require open-ended reasoning as an intermediate step. TypeSafe’s pitch is that, for the subset of tasks that can be specified as structured decisions — classification, routing, extraction, state transitions — a model designed for that shape is fundamentally better suited than a chat model with a JSON output instruction bolted on.

Jev is available in early access via TypeSafe’s developer platform.