ShiftSnap

Inspiration

Hourly workers — warehouse staff, site technicians, healthcare aides — work long, irregular shifts and then face the most dreaded admin of the pay period: the timesheet. Timestamps live in photos, texts, and memory; breaks get forgotten; a missed day means lost pay. Payroll disputes over hours are one of the most common wage complaints in hourly work. ShiftSnap exists so logging a shift is as easy as texting a friend.

What it does

ShiftSnap is a conversational agent that turns natural-language shift notes into accurate, employer-ready timesheets. Log shifts in plain words — "worked yesterday 7am to 6:30pm, break 1 to 1:45" becomes a precise, break-aware time entry. Unpaid breaks are subtracted automatically; net hours, decimal hours, approved hours, and extra hours are computed for you. The agent flags weekdays in the pay period with no entry, so no shift (and no pay) slips through. One click generates a clean, formatted Excel workbook for any pay period, ready to send to a manager or payroll.

How we built it

Built on the Strands Agents SDK. A single Strands Agent drives a tool-calling loop over four typed tools: log_shift (parse and store a shift from natural language), list_shifts (review logged entries), pay_period_summary (totals, approved vs. extra hours, missing days), and generate_timesheet (build the formatted Excel workbook with openpyxl). The agent runs against a pluggable model provider (Bedrock, OpenAI, Anthropic, or a deterministic offline demo model), with a Streamlit chat UI on top and a local JSON shift store — worker data stays private on the worker's own machine. Demo data is fully synthetic.

Challenges we ran into

The hardest part was time itself: parsing "yesterday 7am to 6:30pm" reliably, handling overnight shifts, and making break subtraction unambiguous when workers describe breaks casually. Solved with a dedicated natural date/time parser plus explicit confirmation of parsed entries before storage.

Accomplishments that we're proud of

A full working agent loop — natural language in, verified Excel out — in a single focused project. Break-aware calculations and missing-day detection that match how real hourly payroll works (8 approved hours/day, extras tracked separately). A demo that runs entirely offline, so anyone can try it without API keys.

What we learned

Small, well-scoped tools beat one giant prompt. Four narrow tools with clear schemas made the agent's behavior predictable and its output trustworthy — exactly what you want when the output determines someone's paycheck.

What's next

Photo timestamp ingestion (snap a photo of a time clock, get a logged shift); multi-worker crews and manager approval flows; one-tap export to common payroll formats (CSV, PDF); optional cloud sync via AWS AgentCore for teams.

Built With

• openpyxl
• python
• strands-agents-sdk
• streamlit