Applied AI EngineeringUnit economics for an AI platform, without guessing
How to build a per-customer cost view for a multi-tenant AI platform: pick units, capture cost at the provider, fix attribution, split shared cost.
Applied AI Engineering
Field notes on running LLM platforms, agents, and voice AI in production.
These articles come from building and operating AI platforms for real users: an agent platform’s LLM spend cut by 79%, spend caps at every layer, an autonomous agent confined to an enterprise tenant, a voice receptionist whose tools fail safe, what the Azure Communication Services retirement means for voice agents, routing web and phone enquiries and app support requests into a CRM, and publishing llms.txt for AI assistants. Client names are withheld unless the client agreed to be named.
For the engineering behind the notes, see Applied AI Engineering.
11 articles
Applied AI Engineering
Applied AI EngineeringAdding an AI chat assistant to a static site
How we added an AI chat assistant to a static Hugo site: no new service, no CORS preflight, idle-based lead capture, hard limits and a kill switch.
Applied AI EngineeringScreening vendor pitches out of a web form without losing real leads
Keyword rules for spam held real prospects. How Omnisnia's web-form screening asks an AI which direction a message is going instead, and fails open.
Applied AI EngineeringWhat the Azure Communication Services retirement means for voice AI agents
Azure Communication Services retires as a standalone offering in 2028, and new phone numbers are already restricted. What it means for voice AI agents.
Applied AI EngineeringSpend caps at every layer, and partition pruning as a requirement
How Baseline, an AI intelligence platform, caps LLM spend per collection, tenant and org, and makes BigQuery partition pruning mandatory.
Applied AI EngineeringAdding llms.txt and llms-full.txt to a Hugo site
How we generate llms.txt and llms-full.txt from Hugo content with custom output formats, the gotchas we hit, and why we did it while it is only a proposal.
Applied AI EngineeringSupport tickets from your apps, straight into the CRM
How a customer-facing app, a backend job and an AI receptionist open support cases in Omnisnia CRM with scoped, revocable support tokens.
Applied AI EngineeringOne intake pipeline for our web form and AI receptionist
How we send website enquiries and AI receptionist calls into Omnisnia CRM through one public web-to-lead endpoint, and why a public token is safe there.
Fail-safe tools for a voice AI receptionist
Lessons from building our own AI phone receptionist: tools that degrade safely, caller ID matching, conversation tuning, and live-call verification.
Cutting an agent platform's LLM spend by 79%
How a multi-agent platform's metered model spend fell from about $107 to $22 a day: attribution, three cache layers, model routing and loop caps.
Caging an autonomous AI agent inside an enterprise tenant
The control pattern we design to before an autonomous AI agent touches a Microsoft 365 tenant: identity, network, policy, logging and a kill switch.