The decision layer for your AI
A compact, multimodal model that reads any input and hands you a calibrated decision. Zero generated tokens. Tuned to your domain.
“Transformer T4 at 118% load, ambient 41°C.”
QWhat should the operator do?
It decides in the language of the source.
Your agent calls a giant LLM for every small, uncertain moment in execution, burning hundreds of tokens of reasoning to reach yes, option B, or 3.7 out of 5. Across thousands of calls a day, that means high latency, runaway token bills, and brittle text parsing. Bivia bypasses the fluff, returning instant, structured decisions for those moments.
Great question! Looking at this ticket, the 502 errors right after the deploy suggest an infrastructure issue. That said, checkout is affected, so Billing may want to take a look too. On balance, I'd lean towards Infra, although it's hard to be certain without more context, so it may be worth checking with both teams. Want me to draft a message to them?
Illustrative answer, written to look like a typical chat reply. Tokens estimated at ~4 characters each; time assumes ~50 tokens/s. Both illustrative.
Set the mode for each request. Turn on abstain, and uncertain calls come back flagged.
One decision in a single pass, near-instant. Built for high-volume calls like routing and triage.
think: true.Decides, then challenges the decision before committing. If it is still unresolved, it abstains.
Works with either mode. If the evidence is missing or conflicting, Bivia returns abstained and the reason.
Captured from the live model, shown as returned. Both requests sent with allow_abstain: true; only the mode differs.
One decision model, everywhere you already work. Add it to your coding agent in one click, float it over any app on your screen, or call it straight from your pipeline.
Add Bivia as an MCP server and your agent stops burning a language model on every yes/no, route, or “are we done?” It just decides. Typed, calibrated, and ready to act on.
A floating companion for macOS & Windows. Draw a region or pick a window and decide on anything on your screen. Always on top, always one keystroke away.
Chrome & Edge. Select text, right-click an image, or draw a region on any page, and the decision comes back inline, right where you are.
Docs, Slides and Sheets. Call =BIVIA.DECIDE, =BIVIA.CHOICE or =BIVIA.SCORE right in a cell, and the answer writes back.
The decision UI inside the Articul8 platform, where the documents, the agents and the team already are. Bring a state and a question, see the typed answer.
One HTTP POST from your agent, pipeline or product, with the same typed response every time. Or use the SDK for Python, Node.js, Go or Rust.
Fine-tune Bivia on your policies, SOPs and past decisions, and it decides the way your company does.
What a reasonable organization might do. It leans customer-friendly.
Decides the way Acme does.
Drop Bivia around your existing generative stack. Use your LLM to write and plan. Use Bivia for the bounded decisions around it.
...and yes, there's a lot more where these came from.
Every complex workflow breaks down into a loop of small, typed judgments. Here is how engineering and operations teams wire Bivia into production, with real inputs and real model outputs.
A datacenter cooling loop, decided call by call, with 13 honest abstentions handed back to an operator. One calibrated decision per step, at a volume no language model could touch.
Bivia reads each lot’s defect map, metrology and engineer notes, decides what is safe to run, and hands typed decisions to the scheduler that sequences eight lithography scanners.
Send the state and the shape of the answer. Get back a calibrated, typed decision. Same JSON every time, so there's nothing to parse.
Python
pip install a8-bivia
Node.js
npm install @articul8/bivia
Go
go get github.com/articul8/a8-bivia
Rust
cargo add a8-bivia
curl -fsSL https://articul8.ai/bivia/install | shSign up on the Articul8 platform and create a key.
Get API key →Put it in your environment. The SDKs pick it up automatically.
Send the context and a yes-or-no question. The answer comes back typed, with a calibrated probability.
Check status first. If it's answered, use the value. If it's abstained, Bivia did not have enough information to decide, so send it to a person.
valueThe decision. Act on it.probabilityHow confident Bivia is, from 0 to 1.distributionThe probability of every option.valueAlways null, so there's nothing to act on.abstain_reasonWhy Bivia held back.distributionClose to 50/50: the evidence didn't settle it.Need an answer every time? Send "allow_abstain": false and Bivia always picks one.
Grab a key and make your first call in under five minutes.
Every demo below replays an actual Bivia response, captured from the running model. Watch it choose, score, ground, and say “not sure.”
Same vendor, same question. A general model hedges in prose. A text-only decision model confidently clears it. Bivia abstains and routes to compliance, because the evidence isn’t there.
Video, image regions, verbatim text spans, a French-language PDF, and a raw sensor signal, each decided and grounded. Where a text-only model has to guess, Bivia reads the pixels.
One supplier dispute, nine calls across every modality. The recommendation updates as new evidence lands. Calibration you can watch happen.
An insurance claim payout, five decisions in a row. One flipped call breaks the whole chain. Bivia gives the same right answer every time, while the best general models flip about once in eight runs.
your decisionsp(you'll like it) = high. We'd abstain if we weren't sure.