Sungju Kim / Lab

Agent operations, in public

Previously on AI is the main live demo here — but the point isn't the news feed, it's the pipeline behind it: a scheduled agent collects, filters, deduplicates, summarizes, and publishes static JSON. The site reads that JSON. No LLM runs on your request. Below the app you can see how the agent works, how its token usage is shaped, and how the pipeline evolved.

PoleGraph AI ↗

$ previously-on-ai

{} stack

// loading…

// 01 — agent ops

Agent Ops

What the daily agent does, in one cycle.

01

Collect

Scheduled agent pulls candidate updates from upstream sources on a daily cron.
02

Filter

Drops low-signal items by keyword and source rules before any LLM call.
03

Dedup

Removes near-duplicates by embedding similarity — multiple takes of the same news collapse to one.
04

Summarize

Selected candidates go to an LLM for summary, score, and EN/KO translation.
05

Publish

Writes a single static items.json artifact to the site root.
06

Serve

Visitors hit static HTML + JSON. No LLM call is on the request path.

// 02 — token optimization

Token Optimization sample cycle

Operational metrics for tuning the workflow — not vanity numbers. After deployment these will be replaced by live run logs.

142

Raw items collected

After keyword filter

After dedup

Published items

~280K

Tokens used / cycle

~19K

Tokens per published item

45%

Duplicate removal rate

opus

Model

Prompt version

sample

Last run

Sample cycle · values shown until the deployed agent emits its own run log. · per daily cycle

// 03 — daily usage

Daily usage

Tokens spent per published cycle, one bar per day — the real cost of running the agent over time.

Accumulating — the chart fills in as the daily agent runs.