Narrow jobs, done reliably, with an honest route to a human
We build the assistants, agents and voice products your customers speak into — measured on resolutions rather than messages.
Assistant core
The method
We start from the transaction, not the chat window: which requests the assistant completes end to end, which it triages, and which it hands over. That decision sets the architecture, the tools and the guardrails. Four intents done well beats forty gestured at.
Agents
Agents earn their cost only when the number of steps is unknown at request time. Otherwise one call with structured output and two tools is faster and far easier to debug.
Voice
<800 ms
Speech-to-speech round trip before a caller starts talking over it
Chat
≤2 s
First-token budget past which users re-send and double your load
Discipline
Versioned, reviewed in pull requests and tested against a held-out set before release. Guardrails are specified: what to refuse, what to escalate, what never to state as fact.
Measurement
Containment, resolution verified against the source system, escalation reason, and cost per resolved conversation — the only number that says whether this is worth running.
Models
Orchestration
Voice
Channels
Evaluation
Who we do this for
2–3 weeks
Intent inventory, journey and handover design, and a working prototype on your real content.
Conversation design, prototype, evaluation set, build estimate
8–16 weeks
Full build across the agreed channels, system integration, pilot and a staged production rollout.
Application code, versioned prompts and guardrails, evaluation harness, runbook
Ongoing
Monthly quality reviews, prompt and model upgrades, and new intents as volume shifts.
Quality and unit-cost dashboard, roadmap, quarterly model review
Usually not, at least not at first. If the steps are known in advance — look up an order, check a policy, book a slot — a single call with structured output and two tools is faster, cheaper and easier to evaluate. Agents earn their cost when the number and order of steps cannot be known at request time.
When the process behind it is broken, when nobody owns the content it would answer from, or when the systems it must act in have no usable interface. An assistant in front of a process your own staff cannot complete simply exposes that faster. Fix the process first, or start with the integration work.
Constrain where it gets facts and what it may say. Answers come from retrieval over your own content with the source attached; anything out of scope is refused with a route to a human. Exact claims — a price, a policy limit, an appointment time — come from a system call, not the model.
Send us the constraint you are actually up against — budget, latency, regulator, deadline.
