Elias Andualem

Innovative AI/ML Engineer and Generative AI Specialist with over 6 years of experience in developing cutting-edge AI solutions.

[Elias Andualem @eandualem]

Neither, mostly: it came from taking generation off the hot path. Before, one LLM call per utterance both decided and wrote the movement plan. Now Jev decides in one 0.4 to 0.8 s call and the movement comes from a library of authored plans, so nothing is generated at runtime. Calls are actually more frequent (every fragment), just far cheaper. The LLM still writes a move when the library lacks it, once, and it is learned for next time.

[Elias Andualem @eandualem]

Decision speed in agentic systems, shown on a robot.

In agentic systems, a lot of the work is turning semantic understanding into decisions: which pipeline, act now or wait, stop or continue. Doing that with an LLM works, but each decision costs seconds, and those seconds land in front of everything else.

Jev, the new System One model from TypeSafe AI, is built for that layer. State and typed questions in, calibrated decisions out in well under a second. No text to parse.

Charlie is the demo: a 3D robot in a browser voice call. GPT-Live carries the conversation and only talks. A separate model decides what the body does, and that model was the latency: 3 to 6 seconds between the words and the movement, never in sync.

With Jev as the decision layer, it answers eight typed questions on every fragment of speech in one call, while I am still talking. The body starts about half a second after the words, often before the sentence ends.

Jev is early access.
Avatar Studio: https://t.co/NxNkUyVLgB
Jev: https://t.co/HopK5hqS0O

#Jev #TypeSafeAI #AgenticAI #AIAgents #DecisionModels #OpenSource @typesafeai

[Elias Andualem @eandualem]

The typed-question shape is what makes it usable mid-stream. In a voice call I ask Jev eight questions on every transcript fragment (explicit request? act now? in the library? which entry? energy? duration?) and the answer comes back in one round trip, so a 3D robot's body moves about 0.5 s after the words instead of 3 to 6 s with an LLM deciding. https://t.co/DQ6nMAXGG9

[Elias Andualem @eandualem]

Decision speed in agentic systems, shown on a robot.

In agentic systems, a lot of the work is turning semantic understanding into decisions: which pipeline, act now or wait, stop or continue. Doing that with an LLM works, but each decision costs seconds, and those seconds land in front of everything else.

Jev, the new System One model from TypeSafe AI, is built for that layer. State and typed questions in, calibrated decisions out in well under a second. No text to parse.

Charlie is the demo: a 3D robot in a browser voice call. GPT-Live carries the conversation and only talks. A separate model decides what the body does, and that model was the latency: 3 to 6 seconds between the words and the movement, never in sync.

With Jev as the decision layer, it answers eight typed questions on every fragment of speech in one call, while I am still talking. The body starts about half a second after the words, often before the sentence ends.

Jev is early access.
Avatar Studio: https://t.co/NxNkUyVLgB
Jev: https://t.co/HopK5hqS0O

#Jev #TypeSafeAI #AgenticAI #AIAgents #DecisionModels #OpenSource @typesafeai

[Elias Andualem @eandualem]

@tamarajtran @typesafeai Another one where the LLM was the bottleneck: deciding what a 3D character's body does while the user is still speaking. Eight typed questions per transcript fragment in one Jev call; movement starts about 0.5 s after the words instead of 3 to 6 s. https://t.co/YMfvKVgALl

[Elias Andualem @eandualem]

Decision speed in agentic systems, shown on a robot.

In agentic systems, a lot of the work is turning semantic understanding into decisions: which pipeline, act now or wait, stop or continue. Doing that with an LLM works, but each decision costs seconds, and those seconds land in front of everything else.

Jev, the new System One model from TypeSafe AI, is built for that layer. State and typed questions in, calibrated decisions out in well under a second. No text to parse.

Charlie is the demo: a 3D robot in a browser voice call. GPT-Live carries the conversation and only talks. A separate model decides what the body does, and that model was the latency: 3 to 6 seconds between the words and the movement, never in sync.

With Jev as the decision layer, it answers eight typed questions on every fragment of speech in one call, while I am still talking. The body starts about half a second after the words, often before the sentence ends.

Jev is early access.
Avatar Studio: https://t.co/NxNkUyVLgB
Jev: https://t.co/HopK5hqS0O

#Jev #TypeSafeAI #AgenticAI #AIAgents #DecisionModels #OpenSource @typesafeai

[Elias Andualem @eandualem]

Same pattern one level down: Jev as the router for a robot's body in a live voice call. GPT-Live talks, Jev picks what the body does from a library on every fragment of speech. The old router was an LLM call at 3 to 6 s; now about 0.5 s, before the sentence ends. https://t.co/DQ6nMAXGG9

[Elias Andualem @eandualem]

Decision speed in agentic systems, shown on a robot.

In agentic systems, a lot of the work is turning semantic understanding into decisions: which pipeline, act now or wait, stop or continue. Doing that with an LLM works, but each decision costs seconds, and those seconds land in front of everything else.

Jev, the new System One model from TypeSafe AI, is built for that layer. State and typed questions in, calibrated decisions out in well under a second. No text to parse.

Charlie is the demo: a 3D robot in a browser voice call. GPT-Live carries the conversation and only talks. A separate model decides what the body does, and that model was the latency: 3 to 6 seconds between the words and the movement, never in sync.

With Jev as the decision layer, it answers eight typed questions on every fragment of speech in one call, while I am still talking. The body starts about half a second after the words, often before the sentence ends.

Jev is early access.
Avatar Studio: https://t.co/NxNkUyVLgB
Jev: https://t.co/HopK5hqS0O

#Jev #TypeSafeAI #AgenticAI #AIAgents #DecisionModels #OpenSource @typesafeai

[Elias Andualem @eandualem]

@typesafeai Working integration, three days in: Jev deciding what a 3D robot's body does during a live voice call, eight typed questions per fragment of speech in one call. From 3 to 6 s with an LLM deciding down to about 0.5 s. Video and code: https://t.co/YMfvKVgALl

[Elias Andualem @eandualem]

Decision speed in agentic systems, shown on a robot.

In agentic systems, a lot of the work is turning semantic understanding into decisions: which pipeline, act now or wait, stop or continue. Doing that with an LLM works, but each decision costs seconds, and those seconds land in front of everything else.

Jev, the new System One model from TypeSafe AI, is built for that layer. State and typed questions in, calibrated decisions out in well under a second. No text to parse.

Charlie is the demo: a 3D robot in a browser voice call. GPT-Live carries the conversation and only talks. A separate model decides what the body does, and that model was the latency: 3 to 6 seconds between the words and the movement, never in sync.

With Jev as the decision layer, it answers eight typed questions on every fragment of speech in one call, while I am still talking. The body starts about half a second after the words, often before the sentence ends.

Jev is early access.
Avatar Studio: https://t.co/NxNkUyVLgB
Jev: https://t.co/HopK5hqS0O

#Jev #TypeSafeAI #AgenticAI #AIAgents #DecisionModels #OpenSource @typesafeai

[Elias Andualem @eandualem]

@CompleteSkeptic Put Jev in as the decision layer of a talking 3D robot. It answers eight typed questions on every fragment of speech in one call, and the body reacts about 0.5 s after the words instead of the 3 to 6 s it took with an LLM deciding. Video and code: https://t.co/YMfvKVgALl

[Elias Andualem @eandualem]

Decision speed in agentic systems, shown on a robot.

In agentic systems, a lot of the work is turning semantic understanding into decisions: which pipeline, act now or wait, stop or continue. Doing that with an LLM works, but each decision costs seconds, and those seconds land in front of everything else.

Jev, the new System One model from TypeSafe AI, is built for that layer. State and typed questions in, calibrated decisions out in well under a second. No text to parse.

Charlie is the demo: a 3D robot in a browser voice call. GPT-Live carries the conversation and only talks. A separate model decides what the body does, and that model was the latency: 3 to 6 seconds between the words and the movement, never in sync.

With Jev as the decision layer, it answers eight typed questions on every fragment of speech in one call, while I am still talking. The body starts about half a second after the words, often before the sentence ends.

Jev is early access.
Avatar Studio: https://t.co/NxNkUyVLgB
Jev: https://t.co/HopK5hqS0O

#Jev #TypeSafeAI #AgenticAI #AIAgents #DecisionModels #OpenSource @typesafeai

[Elias Andualem @eandualem]

@MatthewBerman @bot @SpaceXAI My favorite Grok Bot use case:

Testing its limits

[Elias Andualem @eandualem]

@OpenAIDevs My long-running AI coding test: a topological Rubik’s cube synchronizer.

Every model I’d tried had failed. GPT-6 Max got it in one shot, ~40 minutes. Still blown away.

Try it: https://t.co/huw7foovSH

[Elias Andualem @eandualem]

@theo Great thread! Just published on why closing off frontier models (like the new private releases) hurts open research and progress more than it helps: "The Privilege of the Frontier: What We Lose When the Best Models Go Private" https://t.co/tyiKgUJgnO love your thoughts!

[Elias Andualem @eandualem]

https://t.co/GG7LCVucTm

[Elias Andualem @eandualem]

https://t.co/GG7LCVucTm

[Elias Andualem @eandualem]

Check out my latest article: What is the Highest Leverage Use of Agents in 2026? https://t.co/QgtlmhcoPv via @LinkedIn