Summaverick · agentic retrieval
Retrieval that cites its sources and survives production.
I build ServiceNow apps, AI systems, and agents that hold up in production. This is one of them — a research agent that answers from version-matched sources, shows its working as it runs, and publishes the evals it passes and the two it doesn't.
Confident and wrong is the default.
Layered, and version-aware.
The classifier reads the query, then retrieval descends the layers it needs. Live instance query is the third layer and it stays off until a client turns it on.
Injection stops before the model sees it.
Retrieved documents are untrusted input. A poisoned instruction hidden in one is caught, stripped, and logged — the clean evidence continues.
Every claim carries a citation. The evals are public.
Watch it work.
Ask a ServiceNow question. The trace renders stage by stage — classification, the layers consulted, dedup, the injection check, and the verified answer with its sources. An injection attempt is blocked in view, not swept under.
The trace streams live over SSE from the agent — rate-limited, and a query that trips the injection detector is blocked before the model. Try one to see it. (Served statically without the Worker, it falls back to a captured trace.)
Four kinds of work.
ServiceNow store apps, AI transformation, custom agents, and the products that come out of doing all three. Same rule across them: cite the source, show the working, name the constraint.
Selected work.
Client names are withheld by default. The numbers are not.
How working together looks.
Three shapes. Each says what it is, how long it takes, and what you hold at the end — so you can tell whether it fits before we talk.
Start a conversation.
Tell me what you're trying to build. I read every message myself.