Voice and chat AI agents gave conflicting answers on fees and availability at one dental group

·9 views

A four-clinic dental group in Ohio used one vendor and one knowledge base for both its AI phone agent and its web chat agent, yet the two channels answered the same questions differently. An audit of 60 conversations per channel found contradictions on rescheduling fees, insurance acceptance, appointment length, and same-day availability. The most costly case involved the chat agent citing a $25 late-rescheduling fee while the phone agent said there was no charge, which led customers to doubt the stated pricing. The inconsistencies were attributed to system prompts written separately with different priorities, voice-specific brevity that dropped policy qualifiers, and separate feedback channels that never compared transcripts. The suggested fix is a canonical answer table for money- and commitment-related questions, with channel-specific wording that must keep all qualifiers.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

MockMini releases v0.1.0, a small auditable macOS proxy core of about 2,600 lines

MockMini, a compact proxy core written in C and Swift, has reached its first tagged release, v0.1.0, under an Apache-2.0 license. It supports Shadowsocks, VMess, Trojan and VLESS over TCP, TLS or WebSocket, running on its own bounded userspace TCP/IP endpoint. Unlike larger kernels such as mihomo, Clash or sing-box, it avoids vendored proxy kernels, third-party TCP/IP stacks and the Go runtime, so one developer can review the whole codebase. Its packet engine is fuzzed, with about 99% line coverage, and continuous integration runs sanitizers and checks against a pinned Xray release. The project still lacks features such as gRPC, Reality and SACK, and its throughput trails mature kernels, while its maintainers invite reviews and contributions.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Spirula Studio's Vulkan build trains a 100,000-splat Garden scene for Three.js

A test of Spirula Studio's Windows Vulkan build used the official Mip-NeRF 360 Garden sample with supplied camera poses. The 2,000-iteration training run produced a 100,000-splat scene in about 13.6 seconds of total process time. The 24.80 MB PLY file was converted to a 1.95 MB SPZ file and rendered in a local WebGL scene using Spark 2.4.0 and Three.js 0.186.1. The author notes the test did not cover other GPU vendors, new phone captures, or production use in the Digital Garden.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why total queue depth misleads when one Service Bus queue serves many customers

A SaaS platform uses a single shared Service Bus queue to link its cloud product with customers' on-premises systems. The author favors having an outbound-polling agent pull from the broker over inbound HTTP triggers, since the queue absorbs customer outages as backlog instead of failures. Fairness between tenants is handled with per-tenant session IDs, with dedicated queues reserved for high-volume exceptions. The article argues that alerting on aggregate active message count hides a single stuck tenant, as in a case where one customer's credentials failed and their data sat unsynced for days. Instead, it recommends tracking the age of each tenant's oldest unprocessed message, which distinguishes a busy tenant from a broken one.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why a form-validation engine should wait for a second repeat before adding a new rule

A developer received three near-identical requests to block certain characters across different form fields, prompting a decision about whether to keep adding one-off checks or formalize a pattern. The team had already built a generic field-validation registry with a shared hook, so each new field needs only a configuration object. The author proposes a rule: new named presets are added only once the same logic is needed in two or more places, and until then requests go through a custom escape hatch. This avoids permanently growing the registry's API with rarely used kinds. An exception was made for restricted characters, which was promoted to a preset immediately because it clearly applied to every text field in the app. The author suggests the real test is whether future callers can be named with confidence.

0 comments Read more at DEV Community