Restaurant Intelligence Platform

Deecho Improves Customer Service Using Claude

Deecho routes 100% of its production counter audio through a Claude pipeline that decides where each customer transaction begins and ends, and the per-call instrumentation Ravn built let one prompt revision be measured on live traffic, where it cut the cost of the pipeline’s largest stage by 4.5x and the whole pipeline by 37%.

Interested in AI Transformation?

The Challenge

Deecho captures store-floor audio on edge devices in the store to measure service quality. The original approach decided on the device where one customer interaction ended and the next began, using silence as the signal. Those cuts were unreliable: a customer pausing to read the board looked the same as the end of an order, and a slow card authorization looked the same again. Single transactions were split in half, consecutive customers were merged into one record, and because every downstream quality score is computed per interaction, one bad boundary corrupted everything above it. None of the work above the transcript is solvable with simple rules. Distinguishing a mid-order pause from a real goodbye, and telling a generic closer apart from a real upsell attempt, requires reading the exchange the way a person would. Deecho evaluated other models first; they cut conversations in the wrong places, mis-scored service behaviours, and got worse as more context was given to them.

The Solution

Ravn moved transaction segmentation and quality scoring off the device and into a Claude pipeline. Claude reads the conversation and keeps a single customer interaction together across audio chunks. A second pass recovers speaker roles when the transcript is ambiguous. Speaker identity is not used for boundary decisions. Every production model call carries a full per-call cost and prompt-version audit, with cost metered against an effective-dated price book. Before the engagement, some stages were billing without that audit trail; Ravn closed those gaps. Prompt behaviour is pinned by automated golden-fixture tests across the stack.

Results

The headline comparison is controlled: the same pipeline stage, the same model, the same measurement method, one prompt revision apart, with both sides priced at Anthropic list rates from the same effective-dated price book so the only variable is the prompt. Measured from the production audit log. The whole pipeline model cost per five-minute audio chunk fell 37% after the prompt revision. The largest stage fell 4.5x on the same basis. The retry share of that stage’s calls went from 23.6% to 10.9%. Call volume was effectively identical across the two periods, so the saving comes from each call costing less. Separately, routing eligible stages to the batch tier saves 40% of total model spend, and prompt caching now serves 62.4% of input tokens. Adoption is 100% of production audio: every chunk from every commissioned store runs through the pipeline, and there is no manual path. Before the engagement there was no per-call cost attribution; every call is now priced by stage, model, service tier and location. Spend visibility went from a monthly cap hit with no warning, where the code retried instead of degrading, to continuous cost polling with staged alerts and explicit cap-reached handling. In a head-to-head against a competing consumer device on three real transactions using the same audio, Deecho matched or beat the peer on all three and was stronger on speaker tracking.

Tech & Cloud

Claude Sonnet and Claude Haiku via the Claude API (direct). Sonnet handles the heavier transcript and structure work; Haiku handles quality scoring, upsell and service signals, and daily insights. The pipeline uses versioned prompts, the Message Batches API where appropriate, and prompt caching. Ravn built and maintains the pipeline with Claude Code. In production since 2026.

MORE WORK