KorCRM writes my agency's Upwork proposals — three drafts, one survivor, and a gate that ships nothing until the slop score reads zero. The judge is a rival model.

Six views — Today, Hunt, Craft, Deliver, Money, Learn — replace the spreadsheet that used to run an Upwork agency. Mine.

Underneath the views sits a real operating system: a Postgres data spine where eleven automation rules fire on every entity change — connects debited when a proposal goes out, refunds credited on outbid, income mirrored to transactions — each rule atomic inside a single prisma.$transaction, every mutation writing an append-only audit row. Above it sits the machine this dossier is really about: a five-stage AI pipeline that turns a scraped job posting into a scored, gated, human-reviewed proposal.

One honesty note before any drawing: KorCRM was built for one agency — my own. The multi-tenant access layer exists so my team can share it safely, not as evidence of paying customers. What the daily use buys instead is something rarer on a portfolio page: every claim here survived contact with production, because I am the production.

Somewhere tonight a freelancer is pasting proposal number forty into a text box, and the connects are spent whether or not anyone reads it.

The feed refreshed an hour ago. The job looks winnable, so the draft starts — from a blank page, again, because nothing recorded what the winning proposals had in common. The connects ledger lives in one spreadsheet tab, income in another, and the client who hired us in March is a row nobody updated. Most replies never come, and the ones that do arrive carry no lesson: the loop has no memory. Every evening ends the same way — money spent on attention, words spent on instinct, and nothing learned that survives until morning.

FIG. 1 shows the loop as lived. It runs; it never learns; it has no exit.

Four months of production use leaves marks. Here they are, dated and traceable — including the one a portfolio would normally hide.

Read REV D twice; I did. A proxy whose one job is holding secrets sat reachable from the internet because two correct systems — Docker publishing a port, ufw filtering INPUT — composed into a wrong one. The fix took a day. The lesson cost more: on a solo product, you are also the security team, and the only honest response to that is the habit this table demonstrates — find it, root-cause it, bind it, write it down where clients can read it.

What the plate means for anyone running a pipeline of their own: the expensive part of AI writing isn't the writing — it's knowing, mechanically, when the output is allowed to exist, and feeding what happened next back into the grader. That whole apparatus is what you saw in FIG. 2, and it runs for the price of a coffee per proposal.

The plain version, without the drafting frame: I built this because my own evenings looked like FIG. 1, and I was tired of paying for attention with no memory of what earned it.

The pipeline gets the headlines, but the part I'd defend in any interview is duller: 194 recorded decisions, 274 logged gotchas, a revision table that keeps its security incident on the page, and a rework wave where I stopped shipping features for a month to make the architecture tell the truth. A proposal machine without outcome memory is just a faster typewriter — the loop was always the product.

And the honest scope line, one more time: this is a one-agency system, run daily by the person who built it. That's a smaller claim than "SaaS with customers" and a stronger one than most demos — every screw in this drawing has been turned under load, by me, this quarter.