AI workflow automation
Find the repetitive read-and-decide steps in your operation and automate them end to end, with human review where it counts.
Kodiak · open decision model · v0.2
Decisions,
not tokens.
Most of what software asks an LLM to do isn’t writing. It’s deciding: which intent, how urgent, which tool, safe or not. Kodiak answers those questions directly. You send it typed questions and get back calibrated answers in one forward pass, with nothing to parse.
state
Customer: My card was charged twice for order #4411 and I need it fixed before Friday.
-
intent · What does the customer want?choice · p_null 0.02
-
urgency · How urgent is this?score · p_null 0.03no time pressure0.78 [0.57, 0.93]immediate
-
card_brand · Which card brand?choice · p_null 0.91∅ can’t tell from this stateabstain · 91%
Ticket routingTriageTool selectionGuardrails
Policy checksPrompt-injection screeningLead scoringEscalation
Ticket routingTriageTool selectionGuardrails
Policy checksPrompt-injection screeningLead scoringEscalation
01 — The model
A small encoder that does one job very deliberately.
Kodiak is built on Ettin-encoder-1B, an open encoder (ModernBERT architecture) that reads text but can’t write it. On top of that we add a decision layer: a packed request format, a structured attention mask, and choice, score and abstain heads. Every answer stays inside the answer space you defined.
requestPOST /v1/decide
{
"state": ["Customer: My card was charged twice
for order #4411 and I need it fixed
before Friday."],
"questions": [
{ "id": "intent", "type": "choice",
"labels": ["refund or billing fix",
"order status", "cancel order"] },
{ "id": "urgency", "type": "score",
"min": 0, "max": 1 },
{ "id": "card_brand", "type": "choice",
"labels": ["Visa", "Mastercard", "Amex"] }
]
}responseillustrative values
{
"intent": {
"answer": "refund or billing fix",
"confidence": 0.94, "p_null": 0.02 },
"urgency": {
"answer": 0.78,
"interval": [0.57, 0.93] },
"card_brand": {
"answer": null,
"abstain_reason": "unanswerable",
"p_null": 0.91 }
}-
i
Nothing to parse
Kodiak never generates text. A choice is always one of your labels, and a score is always inside your range. No retries for broken JSON.
-
ii
Labels you invent today
Choice labels are defined per request, so a new label set works without retraining. Short, distinct labels work best.
-
iii
An honest “can’t tell”
Any question can come back as unanswerable from this state, with its own probability, so the model doesn’t have to guess.
-
iv
Confidence that means something
Trained with proper scoring rules and evaluated on calibration (ECE, Brier), not only accuracy. When it says 90%, that’s meant to be true.
-
v
Small, fast, cheap
About 38 ms per decision on a GPU, and it also runs on a CPU (a few hundred milliseconds), with no per-token bill.
-
vi
Open, Apache-2.0
Built in the open with every decision and dead end documented. Download the weights on Hugging Face and run it yourself. The hosted API is coming.
02 — Results
Kodiak-v0.2-1B, measured.
An open 1B decision model that matches an 8B LLM on decisions it has never been trained for, about 40× faster.
| Metric | Kodiak-v0.2-1B | v0.2 accuracy mode | Qwen3-8B LLM, our own baseline run |
|---|---|---|---|
| Never-seen tasks, accuracy | 0.689 ± 0.008 | 0.706 | 0.688 |
| Calibration error, never-seen lower is better | 0.085 | 0.062 | 0.293 |
| When it says “can’t tell”, it’s right | 0.88 | 0.905 | – |
| Speed GPU, one request | ~38 ms | ~3× the single model | ~1,500 ms |
Frozen eval set v0.2, choice questions. ‘Never-seen’ = tasks and label sets the model was never trained on. Kodiak-v0.2-1B figures with ± are the mean of three training runs. Accuracy mode averages three models.
New in v0.2
- Checks whether an answer is grounded in its source.
- Picks an assistant’s next step from API specs (call a tool, ask for missing info, or answer directly).
- Verifies claims against a text.
- Judges product relevance.
Known limits
Kodiak has limits worth knowing before you rely on it, including sensitivity to how labels are worded. Read the known limits on the model card ↗
Weights and code are Apache-2.0; the base model is MIT.
03 — How teams use it
Automate the sure ones. Hand off the rest.
Calibrated confidence is what makes this work. Choose a threshold. Decisions above it run automatically, and the ones Kodiak isn’t sure about go to a person or a larger LLM. Drag the gate to see the trade-off.
Simulated queue of 60 decisions, for illustration.
- Automated
- 0
- Escalated
- 0
- Likely errors
- 0
04 — Services
We also build the systems around the model.
Cortex Agent LLC is a small, senior shop. We ship automation that plugs into what you already run, and we tell you when you don’t need AI at all.
Systems & API integration
Connect models and agents to your CRM, ticketing, data stores and internal APIs. We also build MCP servers so your agents can use real tools.
Custom decision models
Teach Kodiak your team’s labels and policies, then run it on your own infrastructure or ours. Your data stays yours.
Architecture & evaluation
Model selection, eval design, cost modelling and the hard question of what should be automated in the first place.
Writing
Notes from the workbench.
05 — Get in touch
The hosted Kodiak API is coming.
Kodiak-v0.2-1B is open and free to download. The hosted API is coming. If you have a decision-heavy workflow and want early access, or you want help automating one now, send a note.
Prefer email? [email protected] · Kodiak on GitHub