PeakShift Ledger

Who it's for

Indie AI developers and small eval/synthetic-data teams running delayable DeepSeek jobs who currently use cron scripts, the DeepSeek pricing page, and spreadsheets to guess cheap run times.

The problem

Payment: delayable jobs run at peak token rates instead of cheaper off-peak rates; Time: developers manually track price windows and reconstruct savings in spreadsheets.

How to build it

CLI plus small web dashboard; user submits a delayable job as a webhook or Docker command with a cost cap, the service schedules it against DeepSeek's published off-peak calendar, sends Slack/email alerts, and exports a CSV/PDF ledger.

How it makes money

Dev teams pay $29/month or 3% of verified savings because one rescheduled eval batch can save over $100, and free cron cannot enforce provider-specific windows, caps, and an audit-ready ledger.

Why it doesn't exist yet

DeepSeek's incentive is selling tokens, not building customer-side tariff schedulers; generic cloud schedulers do not model LLM price windows or produce savings attestations. Indie gap: a thin tariff-aware queue plus cost fence around existing APIs.

First users

First users come from HN and DeepSeek Discord threads showing a side-by-side peak-cost versus off-peak-cost run for one nightly eval batch.

Build size

1 person x 6 weeks; includes CLI, job queue, DeepSeek price-calendar scraper, webhook runner, cap enforcement, CSV/PDF ledger; excludes model hosting, GPU orchestration, and real-time interactive request routing.

Biggest risk

DeepSeek ships native scheduled off-peak jobs or removes time-of-day pricing, eliminating the core savings mechanism.

Conditions for a hit (all 3 required)

• A CLI/web form accepts a delayable job with a maximum token-cost cap and schedules execution only during DeepSeek-published off-peak hours, showing start time, end time, and the price-window source.
• Within 24 hours of job completion it produces a downloadable CSV/PDF ledger showing estimated peak cost, actual off-peak cost, and cited DeepSeek price entries.
• If the estimated cost exceeds the user-set cap, the job is held and an alert is sent before any DeepSeek tokens are spent.

How it's judged (in 6 months)

GitHub repo implementing the shape reaches 1,000 stars or Product Hunt daily top 5(judgment date 2027-02-15)

AI self-confidence 52/100 — self-reported likelihood of meeting the criterion, not a business success rate