RewardRadar

Stop chasing phantom bounties. A source-first multi-agent verifier that checks reward evidence and ranks paid technical opportunities by risk and expected return.

Comment

Inspiration

Open-source contributors and independent builders see large reward totals, but those totals rarely answer the practical question: “If I start this today, what is my real chance of getting paid?” Time-stamped repository capture scripts queried public marketplaces and found closed issues, locked discussions, crowded claims, vague acceptance rules, non-escrowed rewards, and inventories that could not reach a $100 goal. RewardRadar was built to protect a professional’s most limited resource: focused time.

What it does

RewardRadar separates opportunity discovery from verification. Scout collects opportunities, Verifier checks canonical issue or task evidence, Risk Analyst identifies payout and participation risks, and ROI Ranker applies a deterministic expected-value formula. The dashboard exposes the inputs and links to their sources so a contributor can decide whether to pursue, watch, or avoid a task.

The public website is an interactive replay of a clearly dated evidence snapshot, not a continuously refreshed marketplace or a live hosted model. Live-source capture is available through the repository’s separate CLI scripts. Payment probabilities are explicit heuristics, not calibrated guarantees, and advertised rewards are never counted as earned income.

How we built it

The Python agent uses the Strands Agents SDK and GraphBuilder to connect four specialists in an evidence pipeline. Source adapters and canonical verification are exposed as Strands tools; deterministic Python calculates expected value, hourly return, confidence, and verdict. Read-only adapters cover public Opire, Execution Market, Superteam, and GitHub evidence.

The interface is a React/Next dashboard packaged as a static export and deployed with a least-privilege GitHub Actions workflow to GitHub Pages. An owner-private OpenAI Sites preview was used during development.

There are two explicit execution modes. The credential-free DemoModel exercises the actual four-node Strands graph with prescribed fixture tool calls; it is a deterministic test adapter, not an open-ended foundation model. A separate Amazon Bedrock runner is included and prints completion evidence only after Strands returns Status.COMPLETED. The AWS account and Builder ID exist, but Bedrock activation remains incomplete: no completed Bedrock invocation or AgentCore deployment is claimed.

Challenges we ran into

The hardest problem was resisting misleading abundance. An aggregator can label an issue open even when the canonical GitHub issue has been deleted. A high advertised total can hide weak payment evidence, assigned work, substantial competition, or mandatory interviews and trading. We kept arithmetic in deterministic code and made evidence inputs inspectable instead of treating generated conclusions as authoritative.

Accomplishments

  • Implemented a runnable four-node Strands graph and credential-free reproduction path.
  • Verified 22 Python tests from a clean source-archive extraction, plus TypeScript, lint, production build, and dependency checks.
  • Preserved time-stamped September 10 and 11, 2026 source audits.
  • Captured a regression case where an advertised, apparently unclaimed $1,500 reward still looked open on its marketplace but the canonical GitHub issue returned HTTP 410.
  • Added a Superteam verifier that distinguishes the total reward pool from individual awards and surfaces sponsor, public-post, mainnet-trading, and deadline-conflict risks.
  • Published source, a 3:17.74 demo video, and an interactive snapshot. Current public commit: 622ca3ac95afa681ba8167d6f97285972af10673; GitHub Pages build 34812331333 succeeded.

What we learned

Discovery and verification are different jobs. Source quality and uncertainty need to be visible product features. An evidence-backed “avoid” can be the most useful result an agent produces.

What’s next

Add signed marketplace receipts, sponsor histories, region-aware payout compatibility, stronger evaluation data, and opt-in alerts for evidence-backed opportunities. Complete and document a real hosted-model run before making any AWS-runtime claim.

AI assistance, reused components, and limitations

The entrant directed this project with substantial OpenAI Codex assistance for research, source-code generation, tests, documentation, design, and demo-video production. Standard open-source dependencies and their licenses are retained; the repository is MIT-licensed.

The optional CALL-E module uses the MIT-licensed calle-ai 0.7.0 SDK and defaults to a redacted, network-free preview. No live phone call or production CALL-E result is claimed. The read-only Superteam integration performed no mainnet trades and submitted no entry or public social post. No prize or payment is claimed.

Three revised technical articles were published on AWS Builder on September 14, 2026 and are submitted for optional bonus consideration. The source archive retains earlier drafts; no bonus points or prizes have been awarded.

Source and reproduction

Source and instructions: https://github.com/cuentapraces07-ops/rewardradar Public snapshot: https://cuentapraces07-ops.github.io/rewardradar/ Demo video: https://vimeo.com/1226296425

Built With

Share this project:

Updates

Submission history