Everything a filing needs, held ready.
Narrative, video script, and evidence map. A window opening becomes a same-day submission.
Written narrative
Company: The Recording Club, Los Angeles. A membership recording studio with 86 paying members, 79 live Stripe subscriptions, and $44,332 in monthly recurring revenue. Revenue for the first nine months of FY2026 was $606,962.66. July 2026 closed at +$5,059.76 net operating income.
What we built: Gregbot and Claude Grid, an agent operating system that runs the studio’s back office in production. The founder is a working pianist. The system does the work a general manager, a bookkeeper, a release engineer, and an HR coordinator would otherwise do, and it does it inside the real business, against real data, every day.
How the AI is integrated
The system is a fleet of persistent agent sessions, one per project, wired to the founder’s phone through a private tunnel. Voice input is transcribed live on a local model. A text the founder sends to himself is treated as a command. Each agent holds its own memory (832 memory files across the fleet), its own working directory, and its own place in a three-seat doctrine: a dialogue seat that talks and plans, an execution tier of cheaper workers that does bulk and routine work, and a bounded judgment seat that reviews finished artifacts before a person sees them.
Rules are enforced mechanically. Eighteen hook files (1,679 lines) sit between the agents and the world. They block any outbound message that lacks a human approval token, block server creation outside declared cost bounds, block dangerous process kills, and block pushes to the wrong deploy target. When the founder corrects the system, the correction is captured, synthesized into a rule, and where possible converted into a hook rather than prose, because prose rules decay and hooks do not.
The harness is model-agnostic. The seat resolver is a 30-line script that has been repointed six times. A whole-fleet engine switch to a non-Anthropic runtime exists and has been exercised, with all five safety guards ported. Gemini holds a production seat today: it scores every clip in the studio’s short-form video pipeline. Local Whisper does all speech-to-text. The dialogue seat currently runs on Claude; that is a configuration, not an architecture.
What the AI runs
Staff identity. It writes and deploys the authentication server behind every staff login. It minted individual credentials for all eleven studio assistants, verified each against live auth, and shipped a role filter after finding contractors were being read as assistants by role-blind gates.
Payroll. Four server-side refusals block a director’s approve tap when hours, rates, or timestamps are wrong. The AI shipped one of them itself after finding a $960 overpayment caused by clock-outs typed a day late. It reconciled 266 credits totaling $37,636 and proved no other instance existed. The bank rail is built and deliberately unarmed: no token exists, and the endpoint refuses before it touches the database.
Taxes. It rebuilt the filed 2025 partnership return from raw QuickBooks data with no CPA figures as input and converged at 576 matching lines of 583. It found an $800 state payment made twice, five weeks apart, and $4,900 of electronic withdrawal that never drafted because a signed return was never transmitted. It built a $250,773 equipment ledger from 231 invoice PDFs for a state use-tax filing and, when one of its own subagents argued a purchase was exempt, the judge reversed the finding against the client’s interest.
Releases. It runs the App Store pipeline for three shipped apps. When Apple rejected a build for crashing on launch, it reproduced the crash from the shipped binary, wrote a 21-line patch, uploaded, resubmitted seven minutes later, and hardened its own release script so the failure class cannot recur. The build was approved that night.
Membership and finance. Every member is scored 0 to 100 for retention risk nightly. A ledger agent audits credits daily. The month closes from a live Stripe pull before the bookkeeper starts. The July forecast missed actual by $2,327, inside its own published range.
Hiring and operations. 311 applicants were researched and ranked across four departments. Fifteen night-ops templates apply to every event. Session-end surveys fire exactly one staff notification, race-free at the database level.
Business impact
The studio operates with one owner and no back-office staff. Functions that would cost a small business three to four salaries run as software the owner built and supervises. The measurable outcomes are on the record: an overpayment caught and guarded against, a duplicate tax payment recovered, a signed-but-untransmitted return caught before a $3,675-per-month penalty attached, six weeks of silently failed database backups discovered and replaced, and an App Store rejection turned around in a night.
The system audits itself. Every number in this narrative was written down at the time in a memory file, and a second adversarial pass was run on the payroll audit at the founder’s order. That pass reversed several of the first pass’s conclusions and stopped a $5,807.40 double payment. We consider that the most important result: the AI’s errors were caught by the AI’s own process before money moved.
Category
Small Business Services. The harness is packaged as claude-grid-kit for other operators. The demonstration is that a single owner with a domain and an agent OS can run a real company’s operations end to end, with a human hand on every send and every dollar.
Evidence available on request: Stripe dashboard export, monthly P&L with drill-to-transaction, App Store listings, the payroll audit trail, and staff testimonials.
Three-minute demo video · script and shot list
Every shot is a live screen or a live room
-
0:00 to 0:20
Cold open, the phone
-
Shot
Greg at the piano in Studio A, phone in hand. Holds to talk on the Grid app: “Close July from Stripe and tell me the NOI.” Live transcription appears as he speaks.
VOGreg, live. “I’m a pianist. I run a recording studio with 86 members. This is the whole back office.”
-
0:20 to 0:50
The console
-
Shot
Screen capture, grid.tambr.us. Thirteen panes, one per project. The finance pane picks up the request, pulls Stripe, and returns “+$5,059.76 NOI, forecast was +$7,387, $2,327 inside range.”
VO“Thirteen agents, each with its own memory and its own job. A dialogue seat plans, cheaper workers execute, a judgment seat reviews before I see anything.”
-
0:50 to 1:25
It runs payroll
-
Shot
admin.therecording.club, a director taps Approve on a shift with a clock-out dated tomorrow. The server refuses: FUTURE_CLOCK_OUT.
VO“The AI shipped that refusal itself after it found two ex-employees credited exactly 24 hours each, a clock-out typed a day late. $960. It reconciled every credit in the company and proved there were no others.”
ShotThe send-gate approval page. A staged payroll notice waiting for Greg’s click.
VO“It never presses send. It never moves money. I do. That’s enforced by a hook, not a policy.”
-
1:25 to 1:55
It does the taxes
-
Shot
trctax.gregspero.com. The reconstruction board: 576 match / 4 diff / 3 unmapped of 583.
VO“It rebuilt our filed return from the raw books with none of the CPA’s numbers. Then it found we’d paid the state $800 twice and that a signed return had never been transmitted.”
-
1:55 to 2:25
It ships the apps
-
Shot
App Store listing for TRC Staff 2.0.0. Cut to the release log: build 69 rejected 00:12Z, build 70 uploaded 17:55 PT, resubmitted 18:02 PT, approved that night.
VO“Apple rejected a build for crashing on launch. It reproduced the crash from the shipped binary, wrote a 21-line patch, and resubmitted seven minutes after upload. Three apps on the App Store, all released this way.”
-
2:25 to 2:50
Model-agnostic
-
Shot
Terminal. model-policy printed. Then the viral-clips skill file showing VIRAL_MODEL=gemini-2.5-flash. Then the engine switch flipping the fleet to a second runtime.
VO“It’s a harness. Gemini already scores every clip we cut. The seat resolver is 30 lines and has been repointed six times. Any frontier model plugs into this console.”
-
2:50 to 3:00
Close
-
Shot
Back to the piano. Greg plays four bars. Phone buzzes: “July closed. Draft to the bookkeeper is in your drafts.”
CardMono, amber on black. xprize.gregspero.com
Production notes
- Screen captures at 1440×900, 60fps, no cursor smoothing. Terminal font IBM Plex Mono.
- Greg’s audio: lav mic, room tone from Studio A.
- No stock footage, no motion graphics, no music bed until the final four bars (Greg’s own playing).
- Redact member names on any admin screen (view-as-member mode does this).
Submission evidence · the map
What Devpost asks for, where it lives, how long it takes
| Requirement | Status | Source and how | Time |
|---|---|---|---|
| Written narrative, 500 to 1,000 words | Done | Source and howThe written narrative, filed above | Timen/a |
| 3-minute video | Scripted | Source and howThe script and shot list, filed above. Needs Greg on camera and screen captures. | Timehalf a day |
| GitHub repo shared with judges | Ready | Source and howclaude-grid-kit (private). Add testing@devpost.com and judging@hacker.fund as collaborators. | Time5 min |
| Revenue evidence, Stripe dashboard or bank statement with P&L | Ready | Source and howStripe dashboard export plus the P&L dashboard, which drills to the transaction. | Time10 min |
| Expense documentation including marketing spend | Ready | Source and howQuickBooks P&L by month. Google Ads Command Center for ad spend. | Time15 min |
| Product evidence, agent logs, API records, dashboards | Ready | Source and howGrid registry journal (33,081 entries), 22,387 session transcripts, Mission Control census, trctax watchdog log. | Time20 min |
| Customer evidence, contacts and testimonials | To collect | Source and howThree to five members plus two staff, the studio assistants who use the staff app daily. Staff testimonials go out on the ops push lane. | Time2 days |
| Eligibility | OK | Source and howGreg Spero, US resident, over age of majority. | Timen/a |
| Category | Decided | Source and howSmall Business Services. Fallback: Entrepreneurship & Job Creation. | Timen/a |
Round status. Submissions closed August 17, 2026 at 1pm PT. Judging ended September 15. Winners September 25 at 4pm PT. The kit is held ready for the next window or a direct organizer review.