Enterprise messaging and support operations, currently shipping AI into the layer where a contact hits a queue and something has to decide what happens next.
The public builds below all follow one habit: measure first, ship what the number supports, and report the negative results out loud.
contact-triage-model. On 58,638 real customer support conversations, an intake classifier absorbs 22.5% of contacts at 63% precision. Worth $0.45 per contact (95% CI $0.41 to $0.51), $454,781/yr at 1M contacts. Threshold picked on dollars, not F1.
llm-ticket-triage-agent. Support tickets to schema-validated records via Anthropic messages.parse. FERPA scored as recall, not accuracy. Live dashboard.
agent-output-linter. Not what the agent said, but how it got there: did it label itself, did it route a generalized claim to review instead of asserting it as fact, is the queue actually draining. Zero dependencies, 26 tests.
llm-document-scoring-pipeline. LLM does the one interesting call; the pipeline around it is deterministic on purpose. Single-writer lock, size-constrained artifact, docx builder that never re-generates prose.
Reach me on LinkedIn.