Handoff Desk

Inspiration

A completed data job can still be difficult to hand over. Which final file meets the agreed reporting period? What do its columns mean? Were exception rows accounted for? Has a file changed since review? Handoff Desk helps the person responsible for that final review make the evidence visible and keep consequential choices explicit.

What it does

Select CSV candidate versions, state required columns and optionally specify a reporting period. A Strands agent can investigate selected files through bounded tools. The human chooses the governing version and confirms column meanings; deterministic checks review the stated requirements.

The fictional example also checks exact identifier occurrence counts across source, final and exceptions files, preserving leading zeros and repeated identifiers. Uploaded projects do not infer this accounting relationship. A supported review produces a ZIP with required outputs, confirmed definitions, a hash manifest and an unsent delivery email. Changed files invalidate previous evidence. The application does not send email, certify business correctness or establish client acceptance.

How we built it

New Python application code connects Strands Agents to selected-file tools and a deterministic CSV review core. Live investigation runs through AWS Bedrock Mantle using GPT-OSS 20B. The SDK's OpenAI-compatible adapter targets AWS; it does not use a separate OpenAI service account.

The browser records human decisions outside model tool permissions. The public HTTPS demo runs on AWS Lightsail with isolated temporary visitor workspaces, one concurrent investigation and a durable shared 1,000-attempt allowance. Automatic retries are disabled. Call counts, input estimates, output limits and request deadlines bound inference; they are not a dollar spending cap.

Challenges and lessons

Connecting a Mantle credential to the correct AWS endpoint required distinguishing Mantle from Bedrock Runtime. We also separated a human-request deadline from the cumulative model-call allowance, so clarification does not silently renew the quota. Restart cleanup removes abandoned visitor files while preserving the shared allowance. Explicit stop reasons now distinguish limits from unavailable service without returning private provider details.

What we verified

All 69 automated checks passed in the separate public release directory and on the AWS server. They cover conflicting versions, required definitions, exact identifier accounting, changed-file rejection, selected-file restrictions, uploads and exports, request validation, model limits, restart cleanup and overlapping-worker protection.

The recorded demonstration uses an actual live AWS investigation and the real downloaded ZIP, with fictional data. Separate public HTTPS checks verified visitor isolation and upload/export behavior. No real client acceptance or business correctness has been established.

What's next

Pilot a permitted agency handoff and use that evidence to decide which additional file types to support. The free judging prototype is planned through October 8, 2026, with limited shared inference capacity and one-hour inactive visitor retention. No account or payment is required.

Attribution

Built during the hackathon period with AI coding assistance. Uses Strands Agents and its pinned dependencies. Original project source is MIT licensed; dependencies retain their respective licenses. No private DeltaX implementation or client data is included. The video uses disclosed synthesized narration and a clearly labelled inspection of the actual downloaded archive.

Built With

• python
• strands-agents