Finding legitimate paid opportunities online is surprisingly inefficient. Listings go stale, eligibility is unclear, headline prizes can hide poor expected value, and people waste time investigating opportunities that were never realistically actionable.
OpportunityPilot was built to turn that messy research process into an agentic workflow.
What it does
OpportunityPilot takes a user's goal and constraints and evaluates opportunities through evidence-based gates.
It can:
- reject stale or closed opportunities
- check deadline and eligibility constraints
- filter opportunities requiring excessive cost or manual labour
- compare payout, likely speed to cash and expected value
- explain why an opportunity was accepted or rejected
- rank viable opportunities into an execution queue
- identify the exact point where human action is genuinely required
Instead of simply recommending possibilities, the goal is for the agent to progressively narrow the search into an actionable plan.
How we built it
The prototype uses Python with a FastAPI service and an agent layer designed around Google's Agent Development Kit and Gemini.
The deterministic screening engine handles hard constraints and evidence gates, while Gemini is used for reasoning and agent orchestration where judgement is required.
The architecture also includes Google Cloud integration for deployment and persistent evidence storage.
Challenges we ran into
The main challenge was preventing the agent from treating attractive but stale or unverifiable opportunities as valid.
OpportunityPilot therefore separates hard evidence checks from model reasoning. An opportunity must pass objective gates before it can be promoted into the execution queue.
Another challenge was keeping the human out of repetitive work while still recognising actions that legitimately require a person, such as account authentication, identity verification or final submission.
Accomplishments that we're proud of
The current prototype includes:
- an evidence-gated opportunity screening engine
- deterministic acceptance and rejection logic
- ranked execution queues
- an API layer
- Gemini / Google ADK integration
- Google Cloud deployment configuration
- persistent evidence-store support
- automated acceptance tests
The core screening test suite currently passes 11/11 tests.
What we learned
Agentic systems become much more useful when language-model reasoning is combined with explicit state, evidence and deterministic constraints.
The agent should not merely sound confident. It should be able to show why an opportunity remains actionable.
What's next for OpportunityPilot
Next steps include deploying the complete agent on Google Cloud, expanding live opportunity-source connectors, improving expected-value modelling and allowing OpportunityPilot to progress suitable opportunities further before handing unavoidable actions to the user.
Log in or sign up for Devpost to join the conversation.