Turn AI decisions
into reliable systems.
Build with specialist models. Test against real cases, improve failures, and deploy versions you can trust.
“My subscription was charged twice. Can you refund the extra payment?”
See the decisions your app made.
Inspect each call by endpoint and version. See the choice and its scores, and the downstream outcome when your application reports it. Find the misroutes, bad tool picks and missed gates worth improving.
Find the failures.
Iterate from there.
Production decisions give you the cases worth fixing. Revise the decision or fine-tune a specialist. Both go through your saved cases before release.
Turn a mistake into a test case.
Save the input and expected answer, then revise the question, options, model choice or acceptance rules.
Label mistakes. Train a specialist.
Label examples from your own workload. Fine-tune Helm or any Hugging Face model on your data, using your own compute.
Know what changed
before you ship.
Keep production separate from experiments. Check a candidate against your saved cases, then decide what goes live.
Run saved cases.
Compare expected outcomes with the revised decision before release.
Track each release.
Keep immutable versions so you know which configuration served a call.
Choose what goes live.
Publish a tested version separately from the working draft.
Keep a known-good version.
Preserve earlier releases to return to when a new version underperforms.
Choose the model
for your input.
Use Helm 3 when the decision depends on text, images and audio together. Helm 2 stays the text model, and Helm 2 Omni stays available, with the Mechanical specialist on it. Every core model takes the same request.
Text, images and audio together, in one pass. Every question answered with a probability for every option.
Helm 2 Routing
Choose an answering model from the request, candidate capabilities and prices. Evaluate quality and total cost on your workflow.
LLM routing ↗Omni · Mechanical
Which bearing is failing, whether the motor is healthy, if the belt is loose, in one call. 88% by ear on held-out audio, against 30% for Omni alone.
Machine faults by ear ↗Fine-tune on your data.
Start from Helm or any model on Hugging Face. Use your own dataset and compute to train a specialist for the cases your application sees.
Try Saina Studio ↗One context in.
Every question scored.
Send the context once, with every question you need answered. Each gets its own options, scores and next step. The model reads the situation one time, not once per question.
“I paid, but checkout failed. I need this order today.”
Give uncertain cases
a fallback.
Score the options for a text, image or audio input. Set an acceptance threshold: when the top score falls below it, Helm returns no selection so your workflow can review, retry or ask for more context.
“My subscription was charged twice. Can you refund the extra payment?”
Incoming support request · textWhich team should handle this?
single_choiceA specialist, not a generalist.
Helm 3 against the best other model under 1B on each benchmark, with the previous Helm below. Where Helm 3 isn't first, we show it.
Scope. Full run of Decision Index 0.3, every benchmark on its complete set, and the full public vision track. Other models and the Helm 2 rows from the public board of October 26, 2026. Self-run with the published harness; Helm 3 is not on the public board yet.
Limits. Weakest on knowledge-heavy questions. GPQA Diamond 30.6 and HLE 12.4 are at or near chance, MMLU-Pro is 30.3 and PhishNChips 51.5 is near chance. Not the model for those.
Common use cases.
Same request format.
Each step has a page with the exact request, what comes back, where to set the confidence threshold and how to wire it in.
Route incoming messages.
A ticket arrives; Helm picks the queue. One n8n node or one HTTP request.
Helm 2 · Ticket triage ↗02 / TEXT + IMAGETag and prioritise work.
An email lands in a shared inbox; Helm scores urgency and team, screenshot included.
Helm 2 Omni · Inbox routing ↗03 / MODEL SELECTIONPick the model, or gate the output.
A request enters the pipeline; Routing picks the model, or gates the action first.
Helm 2 · Routing · LLM routing ↗04 / IMAGECheck receipts against claims.
A receipt photo arrives with a claim; Omni scores whether they match.
Helm 2 Omni · Receipts ↗05 / AUDIOMonitor equipment by sound.
Ten seconds from a contact mic; the Mechanical specialist scores which fault it hears.
Omni · Mechanical ↗06 / AUDIOScreen calls for synthetic speech.
A call comes in; Omni scores whether the voice is synthesized. Borderline goes to a person.
Helm 2 Omni · Voice screening ↗Choose where your
decisions run.
Deploy Studio where your team works, or call the models directly. The inference provider determines where hosted Studio inputs are processed. Studio also works from compatible development tools through its MCP integration.
curl --fail-with-body "$SAINA_STUDIO_ORIGIN/v1/studio/endpoints/support-triage/decide" \
-H "Authorization: Bearer $SAINA_API_KEY" \
-H 'Content-Type: application/json' \
--data '{"version":"production",
"state":{"message":"My subscription renewed, but I see two identical charges."}}'What do you want
to do next?
Pick the task you're working on. These paths link to the relevant product details until dedicated step-by-step guides are available.
Practical questions.
API details, data policies and installation instructions are in the linked documentation.
Create an account, open Saina Studio and create an API key. The free plan grants one credit on each UTC day you use it. Copy the quick-start request, set your key and inspect answers and usage. Use decision mode when you want threshold-based acceptance.
Start with the decision you keep patching.
Try a routing, gating or verification step on your own cases. Add Studio when you need shared tests, controlled releases and outcome tracking.