Build the decision layer for agents with TIA

Many small decisions

Agents and automated workflows ask all day: which team owns this ticket, is this order risky, is this message a scam.

General LLMs are slow and costly per decision

Each decision is a round of text generation, billed for input and output tokens. Multiplied by thousands a day, the time and the cost add up.

Free text can be invalid

A written answer can name the wrong label, break the format, or give probabilities that do not add up to 1. The code downstream has to check and repair it.

See how to define a decision like a tool call

Every number names its split, its source and its date. Where TIA is still weak, we say so.

• PlaygroundCalls the model currently served through the System One API (revision eff27e8). No canned answers, no account needed. Candidate TIA 1.5 has not been deployed to the API.Open the Playground
• acc 0.760 · KL 0.138TIA 1.3 mini on the internal typed-decisions dev set (930 decisions, carved from train), measured by TIA on Sep 30, 2026. These are stored report results, not current API measurements. Dev results are not directly comparable to competitors’ test results; the pre-registered typed-decisions test read has not happened.See the benchmark and its limits
• 69 msp50 of TIA 1.0 (the build measured in the technical report), server-side (model time), one request of 3 questions, 98 tokens, sequential, n = 200; p95 81 ms. Measured on AMD Ryzen 9 9950X3D 16-Core Processor, 32 vCPU, Sep 26, 2026. Not end-to-end latency from a browser; the build behind the API has no published server-side latency yet.
• 500 đper 1M input tokens, output is free; 100 million free input tokens every month for every account. Current list price, paid in VND through PayOS.See pricing
• JevBench 40.3%TIA 1.3 mini, a historical TIA build, scored 29/72 on the JevBench public "original" set with the unmodified JevBench v1.4.2.2 harness, run by TIA on Sep 30, 2026. On the same 72 items, JevBench v1 (Sep 19, 2026) published 93.1% to 98.6% for 7 of its 8 systems. This is a comparison conducted by TIA, not an independent evaluation or official ranking. TIA still has an accuracy gap; partnerships aim to improve and verify this capability.JevBench v1 results

Nguyễn Phương Anh Tú, founder and Admin of 8 Sync, AI Engineer at CloudGO, who directly trained and launched the TIA model family.

CloudPro CRM

owns AI feature delivery on a CRM used by 2,500+ businesses

Teaching

1,000+ students taught across Python, Web, AI/ML and software engineering

Research

two peer-reviewed papers in the JTE journal, as first author

NVIDIA Inception

member since August 2026, as founder of 8 Sync

For partners, investors and early sponsors. Which of these fits you is something we settle by email.

• Early access to new TIA versionsUse new versions before they reach the public API.
• Dedicated capacityCompute set aside for your traffic, separate from the shared public API.
• Deployment in your infrastructureTIA running inside your own company infrastructure, by separate agreement. There is no public build to install yourself.
• Decision models built with youBuilt on your own data, under a non-disclosure agreement (NDA).
• Sponsor acknowledgementEarly sponsors are acknowledged by name or logo on the TIA website.
• Research collaborationMeasure and write together about typed decision models and probability calibration.

This page states no prices, equity terms or SLA commitments. Every detail is agreed one to one.