Clearweek

Know which client payment needs attention before payroll.

Comment

Inspiration

Payroll has a fixed date. Client payments can slip. For a small service agency, that mismatch can put salaries at risk even when the business has enough work. Clearweek helps an owner answer: which invoice needs attention, by when, and would earlier payment cover the gap?

What it does

Import receivables and payroll/vendor CSVs, opening cash, a reserve floor, and an as-of date. A background agent investigates the next 28 days and proposes a payment request with traceable calculations.

In the synthetic Juniper case, a $12,000 invoice is expected September 25, after September 18 payroll. Baseline minimum cash is -$10,000, with a $4,000 payroll shortfall. Receiving that invoice before payroll would raise minimum cash to $2,000, conditional on the customer actually paying.

The owner inspects evidence, edits the requested date, recalculates, and approves or rejects the plan. Approval saves the reviewed version; HTML/CSV exports support the owner's follow-up. A fresh import starts another investigation and marks earlier decisions outdated.

How we built it

Real Strands Agents SDK 1.55.1 runs Amazon Bedrock Nova Pro by default. The agent chooses invoices and candidate receipt dates to investigate, requests scenarios, compares results, and prepares a proposal. Its six bounded tools inspect cash and receivables, calculate the baseline and scenarios, compare scenarios, and prepare the decision.

The agent chooses which supported scenarios to test and runs the inspect, calculate, and compare steps in the background. Deterministic application code computes every financial figure in integer USD cents and validates the finding. Scenarios change expected receipt dates, never invoice amounts or payroll commitments.

FastAPI, SQLite, Jinja2, and vanilla JavaScript provide background jobs, source evidence, human review, and versioned exports. Credentials stay server-side. AgentCore is not used.

Challenges and what we learned

Recalculating a date can make the original explanation stale. We preserve that explanation as history and show current reviewed figures separately. Tool activity also exposes failures and recovery: the captured paid-source recheck recovered from a rejected tool attempt.

Verification and AI assistance

Automated checks cover calculations, agent tools, API contracts, revisions, exports, and UI behavior, alongside Ruff and secrets scanning. A real Bedrock browser run exercised import through investigation, date edit, approval, export, and paid-source recheck.

Codex assisted implementation, review, regression coverage, browser verification, and submission materials. The 3:15 demo is a guided screenshot walkthrough captured from that real run with synthetic data and AI-generated narration.

Limitations and what's next

Inputs are manual CSV snapshots. There are no bank, accounting, payroll, email, or payment integrations. Clearweek never sends messages, moves money, or marks invoices paid. It supports USD, 28 days, and session-scoped data. Customer demand and outcomes remain unvalidated; agency-owner testing comes next.

Source and testing

MIT source and judge runbook: https://github.com/zaeem-rafiq/clearweek-agents-for-humans

Watch the 3:15 walkthrough: https://youtu.be/jDff3JR6gWQ

After setup, choose Juniper v2 — delayed invoice, click Import sample & investigate, change the requested date to September 16, 2026, recalculate, approve, and open the HTML packet. Then import Juniper v3 — paid to see the fresh check.

The free local setup uses Qwen2.5 7B through llama.cpp with real Strands tools, requiring about 5 GB of downloads. Verified hardware: Apple M4 Mac, 16 GB RAM; other hardware is unverified. No cloud key is needed. The demo uses Bedrock; there is no hosted deployment.

Free setup: https://github.com/zaeem-rafiq/clearweek-agents-for-humans/blob/main/docs/submission/free-local-judge-setup.md

Built With

Share this project:

Updates

Submission history