Learn which actions work
Cleared payments and renewals connect conversations to results. Reinforcement learning rewards the actions that led to those outcomes
Learn
Models that learn from your results
Post-train open-weight models on your repayments, renewals and claims. Training runs in your environment, every release is checked against your rubric, and you own the weights
Live modelv7v8 queued
Verify a live call
Maya Patel
+1 (213) •••-8034Verified payment
$210received
Loan systemGuardrails
3of 3 held
Your rulesTrain on what happens after the call
You control every release
A candidate replaces the live model only after it passes your rubric. Your team approves the release
Improve your model, release by release
Compare quality and response latency before a candidate moves into production
Release performance
v7
- Call score
- 94/100
- p95 response
- 380ms
Call score
Coinbase Engineering
Post-trained 9B fraud model
Read the result- +9.6pp
- F1 compared with a frontier model
- < $100
- Total training cost on a single GPU
- −55%
- Median inference latency after self-hosting
From call history to a tested release
Remove personal data, link verified outcomes, train a candidate, and check it against the live model
- PrepareCollect · Redact · Curate
- Trainv8-rc1
- Evaluate6 release checks
- ReleaseCanary · ServeNot promoted
Release checks
| Your rubric | Livev7 | Candidatev8-rc1 | Change |
|---|---|---|---|
| Compliance | 100 | 100Passed | 0 |
| Required disclosures | 100 | 97Failed | −3 |
| Resolution | 74 | 77Passed | +3 |
| Right-party verification | 98.2% | 98.6%Passed | +0.4 |
| Latency p95 | 420 ms | 395 msPassed | −25 ms |
| Regression suite | 412/412 | 412/412Passed | 0 |
Release blocked
Required disclosures fell by 3 points
What we post-train
High-volume decisions with a clear, measurable result
Next best action
Choose a payment plan, a payment link or a handoff to your team
Risk and fraud detection
Score liveness and synthetic-voice risk, and identify AI assistants
Compliance checks
Check identity and required disclosures before account details are shared
Promise and lapse risk
Estimate whether a promise to pay will hold or a policy will lapse
Your data, weights and keys stay in your account
Your cloud
Your keys- Channels
- Know
- Agents
- Measure
Your model
Post-trained weights
Your systems
Private connectors
Conversation data
Recordings · Transcripts · Training sets
Customer data stays inside
Actioneer
Control plane
Signed inbound
- Software updates
- Base model weights
Optional outbound
Aggregate health metrics
No content or PIICommon questions
Who owns the post-trained weights?
You do. Models trained on your data belong to you, run in your environment and are never used for another customer.
Which open models do you start from?
Open-weight model families between 8B and 70B parameters, chosen per task for quality, latency and cost. We benchmark candidates on your rubric before training.
How do you stop the model gaming the reward?
Rewards come from verified events in your systems, such as a payment received, and from rubric scores. Each candidate must pass the full rubric, including compliance rules, before release, so a model that raises payments by breaking a rule fails the gate.
Do we need an ML team?
No. We run the pipeline inside your account: curation, redaction, reward design, training, evaluation gating, canary release and rollback. Your team approves each release.
Do open models replace frontier models?
Not always. Post-training targets the narrow, high-volume decisions inside each conversation. Open-ended dialogue can use a base model or an approved frontier model, under the same policies.
What does it cost to train?
Coinbase reports under $100 for a 9B model on one GPU. We size your run in the pilot.
Test it on your own call history
Bring a batch of past calls. We’ll score the baseline, train a candidate in your environment, and compare the results