Inspiration
What it does
How we built it
Challenges we ran into
Accomplishments that we're proud of
What we learned
What's next for Impact Loom
Inspiration
Small nonprofits spend scarce staff time reconstructing grant reports from attendance exports, facilitator logs, surveys, and spreadsheets. Missing proof can threaten a renewal, while copying an unconsented record can create a privacy problem.
What it does
Impact Loom aggregates only consented evidence, checks quantitative grant requirements, drafts a cited report, and creates follow-up requests for missing evidence. Every external action enters a pending queue and is blocked until a human approves that exact draft.
How we built it
We built a Python 3.12 application with the Strands Agents SDK 1.53.0, FastAPI, Uvicorn, a deterministic evidence workflow, and an append-only SHA-256 audit chain. The Strands agent exposes only two bounded tools: inspect_grant_case and list_pending_approvals. It has no arbitrary shell, filesystem, email, or network tool. Amazon Bedrock is the intended model provider once the Builder ID and sandbox access are complete.
Challenges
The main challenge was separating model reasoning from consequential state changes. Consent filtering, totals, state transitions, and audit verification therefore stay deterministic. The system rejects execution before approval and fails closed if its audit chain is altered.
Accomplishments
The synthetic demo verifies 37 sessions and 84 learners, excludes an unconsented record, finds two reporting gaps, and demonstrates approval-gated execution. Nine automated tests cover consent, gaps, approval blocking, single execution, tamper detection, duplicate and non-finite evidence rejection, Markdown escaping, fingerprints, and the audit API.
What we learned
Human approval is most useful when it is bound to an exact content fingerprint. A generic approval flag is not enough if the draft can change afterward.
What's next
Connect the same bounded workflow to Bedrock and carefully scoped AWS storage and messaging services, while preserving the audit trail and human approval boundary.
Impact Loom was created from scratch during the hackathon. OpenAI Codex assisted development; all generated code was reviewed and tested. The project is MIT licensed.
Log in or sign up for Devpost to join the conversation.