Skip to content

Work

The work

Five systems, five different ways they could have gone wrong. Each case study says what it does, how it is built, and what it had to get right.

Case studies

Diagram of the intake funnel feeding a confidence judge that routes drafts to human review
In productionMar–Jul 2026

Support Inbox Agent

Every support email gets a first draft in seconds. A person still decides what sends.

Node/TypeScriptBraintrust evalsLLM-as-judgeMCP
Architecture diagram showing Slack, a Node app on Railway, and parallel help-center and wiki retrieval
In daily internal useOct 2025–Jun 2026

Internal Knowledge Assistant

1,090 questions answered in Slack, each one cited back to the documentation.

RAGMCPSlack BoltNode.js
Diagram of three flows — token minting, scheduled sync through a cleaner, and server-rendered page loads
In productionJul 2026

Customer Status Pages

Every customer gets a live page for their tickets. A model writes the updates; five checks decide whether they publish.

Next.jsSalesforce Bulk APIGuardrailsEvals
Diagram showing an LLM classifier feeding a template registry and deterministic PII guard, never writing the reply body
Complete, merged into the agent platformMar–Jul 2026

Payments Operations Agent

Ten minutes of manual work per payment notice, automated — without the model writing a word of it.

TypeScriptStructured outputFail-closedPII guards
Diagram of the fail-soft orchestrator fanning out to nine service calls with audit rows for failures
In productionMar–Jul 2026

Customer Onboarding Platform

New customers set up their own account instead of waiting on a specialist.

Next.jsTypeScriptSalesforceSupabase

Got something that has to work?

Tell me what you are building and where you think it breaks.

Get in touch