SaaS · 2024
CustomGPT Researcher.
Deep research, with receipts.
presentsAdmit one
Act I
The challenge
The research agent runs for twenty to thirty minutes and reports back as a stream of free-form text, no structure, no schema. My job was to make that wait feel alive as a live seven-step run, and to turn that raw stream into a clean, fully cited report a reader can trust.
Act II
What I built
The production frontend and server-action BFF for CustomGPT.ai’s long-running research pipeline. The crawl and generation are an external backend; I submit jobs, poll progress, live-render a seven-step run, and reconstruct a fully cited report by parsing a free-text progress stream client-side.
Intermission, the demo
The product, playing itself
the front end for CustomGPT.ai’s research agent, I turn its free-text progress stream into a live 7-step run and a cited report
The cast
What actually makes it work
Read the card, then flip it to watch the mechanism run.
0x01
Report from unstructured logs
The backend streams free text; I detect section markers, regex-extract {name, url} citations, and diff against previous messages to rebuild a structured, cited report on the client.
how it works
Report from unstructured logs
0x02
Multi-task polling manager
A per-task timeout map with resumable tracking, unmount cleanup, and cascade refresh keeps several 20–30-minute jobs alive and correct across navigation.
how it works
Multi-task polling manager
0x03
Dockerized k6 load suite
Smoke, stress (200 VU), and soak profiles with env-tunable thresholds, so the polling layer is proven under load, not just in a happy path.
how it works