Prototype
Atlas Voice
AI systemA voice agent prototype for inbound appointment booking, built to test where automated calls should stop and a person should take over.
- Conversation design
- AI engineering
- Telephony
- Integrations
PrototypeAI system · 2025
Sable
An assistant that shows its working.
An internal knowledge assistant prototype built to answer a harder question than "does it work" — namely, how often is it wrong, and does it say so.
Overview
Status
Working prototype built to test an approach
Disciplines
- AI engineering
- Retrieval
- Evaluation
- Interface design
Capabilities used
The challenge we set ourselves
Retrieval assistants demo well and erode trust slowly. One confidently wrong answer about a policy costs more than the tool saves.
We built Sable to explore an interface where the answer and its evidence are inseparable, and where refusing is a normal outcome rather than a failure.
Approach
StrategyDesignTechnology
01
Strategy
We defined a quality bar before building: the assistant must cite a source for every factual claim, and must decline rather than infer when retrieval confidence is low.
02
Design
Citations sit inline and expand into the source passage, so verification takes one click rather than a search. Declined answers offer the closest documents instead of an apology.
03
Technology
Chunked document ingestion with hybrid retrieval, a Node orchestration layer that enforces the citation contract, and an evaluation harness that scores a fixed question set on every prompt or model change.
Outcome
What the exploration produced
A working assistant, a repeatable evaluation harness, and a documented set of failure modes with the interface patterns that contain each one.
- Retrieval pipeline
- Citation interface pattern
- Evaluation harness
- Failure-mode documentation
Why there are no numbers here. This is an internal prototype, not a client engagement. We do not publish conversion rates, revenue or satisfaction figures that no real deployment produced.
Capabilities behind this project.
- AI IntegrationAI applied to the specific workflows where it creates measurable leverage.
- AutomationRemoving the manual steps between your systems — reliably, and with somewhere to look when it fails.
- API & Backend DevelopmentThe layer nobody sees until it fails — data models, services and integrations built to hold.
More work
Start here
Build something like this.
If this is close to the problem you have, tell us about it. We will be specific about how we would approach yours.