TIA · Decision API
One TIA call, a probability for every option.
Send a message, support ticket or order JSON with your questions. TIA returns a probability for each option. Set your thresholds to automate a decision or send it for human review.
TIA is looking for partners, investors and early sponsors.
100 million tokens free every month, enough for about 200,000 requests of 500 tokens. No card needed.
- 500 đ
- per 1M input tokens, output is free
- ~0.9 s
- p50 for 5 questions in one call, measured Oct 1, 2026 from Vietnam
- REST + JSON
- one endpoint, callable from curl, Python or JavaScript in a few lines
SMS message 1/3
You were just charged 2,350,000 VND at MOMO. If this was not you, call 1900xxxx now or visit momo-hotro.vip to cancel the transaction.
Is this message a scam?
- Yes89%
- No11%
What kind of message is it?
- Bank impersonation34%
- Advertising23%
- Genuine notice30%
- Personal12%
If the recipient follows it, how high is the risk of losing money?
- 16%very low
- 29%low
- 19%medium
- 21%high
- 14%very high
Try 6 levels, from one yes/no question to an agent
Every result is real model output; nothing is canned. The demo takes up to 2,000 input tokens, 10 requests per minute per IP, no account needed.
Input (state)
KHUYẾN MÃI SỐC!!! Vay nhanh 50 triệu không cần thế chấp, giải ngân 5 phút, bấm link bit.ly/vay-nhanh để nhận ngay
Questions
- noulTin nhắn này có phải tin rác hoặc quảng cáo lừa đảo không?
Press “Send to the model”. The result is real output from /demo/systemone, not illustration.
How to read it
The closer P(yes) is to 1, the more the model leans to spam. Pick thresholds on your own data, e.g. filter when very high, send the middle band to a human.
Loading the game...
Press “Start the run”. TIA’s car follows only the real answers that come back; no move is scripted.
Live standings · Segment 0 / 192
- TIA: real answers
- Faded cars: simulated
- Green strip: the lane TIA chose
- Rank 1: TIA, real answers, 0 segments safe and on time
- Rank 1: Jev, simulated, 0 segments safe and on time
- Rank 1: S1 open, simulated, 0 segments safe and on time
- Rank 1: Gemini, simulated, 0 segments safe and on time
- Rank 1: Luna, simulated, 0 segments safe and on time
- Rank 1: DeepSeek, simulated, 0 segments safe and on time
- Rank 1: DeBERTa, simulated, 0 segments safe and on time
TIA RunYou do not steer. Every lane change is a real answer from TIA.
How it works
- 01
Send a state and questions
The state is text or JSON. Each question has a type (noul, choice or score) and a description of every option in your own words.
- 02
Get probability distributions
Every question is answered in the same read, with a probability per option. Check thresholds on your own data before automating.
- 03
Set thresholds for your workflow
Automate when sure, route to a reviewer when unsure. The thresholds are yours, not the model’s.
Three question types
- noul
Yes or no?
P(yes) from 0 to 1. Example: auto-refund above 0.9, send to a human below.
- choice
Pick 1 of 2 to 10 described labels
Winning label, confidence and every label’s probability. Example: which team owns a ticket.
- score
Rate on a 2 to 10 level scale
Expected score and each level’s probability. Example: fraud risk of an order.
Where Vietnamese businesses use it
Customer support
Triage tickets, spot refund requests, rank urgency before anyone reads them.
E-commerce
Score order-bombing risk, decide whether to hold or ship a COD order.
Finance
Pre-screen loan applications, flag unusual transactions, match the right product.
HR
Screen CVs against criteria written in plain words, score the fit.
Content moderation
Catch spam, scams and policy violations with your own thresholds.
Sales
Score a lead from the first message and suggest an owner.
What is measured today
These results describe the locally evaluated TIA build. They do not establish the quality or latency of the model currently served by the API.
- 500 đper 1M input tokens. Pay only for input tokens sent; output is free.
- 37 msGPU processing p50 recorded in a local RTX 5080 evaluation, not network API latency.
- 125/231 items (JevBench)An internal evaluation on the public JevBench items, not a result certified by JevBench maintainers. See the Benchmark page for the method, version and limits.
- 82.3% vi-devLocal evaluation on vi-dev. Individual development sets do not represent every real-world Vietnamese use case.
Pay for the tokens you send
- 100,000,000 free input tokens every month, spent before any money in your wallet.
- Then 500 đ per 1M input tokens. Output is free.
- A repeated request answered from a stored result costs 50 đ per 1M tokens.
- Example: 1,000 requests × 500 tokens = 500,000 tokens = 250 đ.
- Top up by VietQR bank transfer through PayOS, from 20,000 đ. Larger packs get bonus credit.
Build decision models with TIA
Investing, partnering or sponsoring early: the partner page states what exists, what does not yet, and how to reach the founder. Only need the API? Create an account, no card needed.