About the Project BountyPilot AI is an autonomous workflow agent that helps creators discover, evaluate, and execute opportunities — from content bounties and creator campaigns to hackathons, grants, competitions, and innovation programs.
Creators often spend more time searching, researching, and organizing than actually building. BountyPilot AI was built to automate that entire upstream process: analyzing opportunities, extracting requirements, generating research briefs, building execution plans, tracking submissions, and guiding users from discovery to outcome.
The goal is simple: less time managing, more time creating.
Inspiration The idea started from a personal challenge — going from $$0$ to $$1000$ through content opportunities.
I built a Google Sheets system to manually track content bounties, manage scripts, record submissions, and log earnings. It worked. But I quickly realized the real problem wasn't organization — it was the sheer volume of manual cognitive work required before a single piece of content was ever created.
Every opportunity followed the same repetitive loop:
$$\text{Find} \to \text{Read} \to \text{Research} \to \text{Decide} \to \text{Plan} \to \text{Submit} \to \text{Track}$$
And every step was manual. As I explored more platforms — hackathons, grants, creator campaigns, innovation programs — I found the same friction everywhere. The workflow was universal. The automation wasn't.
While competing in the Qwen Cloud Global AI Hackathon, something clicked: I could build the tool I wished I had at the start. That realization became BountyPilot AI.
How We Built It BountyPilot AI runs on a full-stack architecture combining web extraction, AI reasoning, product analytics, and human-in-the-loop checkpoints:
Infrastructure: Replit · Node.js · Express · PostgreSQL · Drizzle ORM · React · Vite · Tailwind CSS
AI Layer — Qwen (Alibaba Cloud): Qwen powers the core intelligence of the platform. Every time a user submits an opportunity URL, Qwen handles requirement extraction, opportunity scoring, research brief generation, and execution plan synthesis. The model is configured via QWEN_API_KEY, QWEN_MODEL, and QWEN_BASE_URL, with a rule-based fallback for development environments without API access.
Analytics Layer — Novus: Novus is integrated as the product analytics engine, tracking how users move through the platform — which features they engage with, where drop-off occurs, and how the workflow performs across different creator segments. This data drives product decisions and helps surface which parts of the automation are delivering real value versus creating unnecessary friction.
The core workflow follows:
$$\text{Discover} \xrightarrow{\text{scrape}} \text{Analyze} \xrightarrow{\text{Qwen}} \text{Approve} \xrightarrow{\text{human}} \text{Research} \to \text{Build} \to \text{Submit} \xrightarrow{\text{Novus}} \text{Track & Improve}$$
Users paste an opportunity link and receive structured analysis, a research brief, an execution plan, and a live submission tracker — all inside a single dashboard. The system also supports a Telegram alert layer so creators receive real-time notifications for newly matched opportunities without ever opening the app.
Challenges We Faced Building through vibecoding on Replit was the foundation of development. Iteration was fast, but daily credit ceilings sometimes halted momentum mid-feature. The solution was ruthless prioritization — ship the critical path, queue the rest.
Waiting for Qwen API access through the hackathon program meant building on mock responses for a significant portion of early development. The architecture had to be API-agnostic from day one, with Qwen slotting in cleanly once credentials arrived. This constraint actually made the system more robust — the mock layer became a reliable fallback for offline development.
Reliable data extraction proved harder than expected. Many opportunity platforms use dynamic rendering, inconsistent HTML structures, and incomplete metadata. Extracting rewards, deadlines, and eligibility requirements reliably required multiple parsing strategies and graceful degradation when data was missing.
The most intellectually interesting challenge was teaching Qwen to identify the correct subject of an opportunity. Early versions would generate research briefs about the platform hosting a bounty rather than the project sponsoring it — a subtle but critical distinction. A Solidity bounty listed on a marketplace would produce a brief about the marketplace itself instead of the protocol behind the work.
Formally, if we define:
$$S = \text{true subject}, \quad P = \text{hosting platform}, \quad C = \text{raw page content}$$
The naive model learned $\hat{S} \approx P$ when $P$ was more prominent in $C$. The fix introduced a structured extraction prompt that enforces:
$$\hat{S} = \underset{s \in \text{entities}(C)}{\arg\max} \ \text{relevance}(s,\ \text{task_requirements})$$
Novus analytics were essential here — by tracking which briefs users rejected or edited before using, we could identify patterns in where Qwen was getting the subject wrong and iterate on the prompts with real behavioral signal rather than guesswork.
Each challenge directly improved the architecture and made the product more reliable under real-world conditions.
What We Learned Creators don't need more dashboards. They need systems that reduce friction.
The biggest lesson is that AI product quality isn't just a function of model intelligence. It's a function of four things working together:
$$\text{Quality} = f(\underbrace{M}{\text{model}},\ \underbrace{D}{\text{data}},\ \underbrace{W}{\text{workflow}},\ \underbrace{H}{\text{human oversight}})$$
Qwen provided strong reasoning out of the box, but the real leverage came from engineering the right context into every prompt — structured opportunity metadata, creator profile signals, platform-specific extraction rules. Novus then closed the feedback loop, translating user behavior into concrete signals about where the workflow was working and where it needed refinement.
We also learned that the best products often start as the solution to a problem the builder lives inside. What began as a personal $$0 \to $1000$ challenge grew into a platform designed for every creator who has ever lost hours to opportunity research instead of opportunity execution.
Long-Term Vision The vision is to build the operating system for opportunity-driven creators.
Whether someone is pursuing content bounties, hackathons, grants, competitions, or creator campaigns, BountyPilot AI should take them from opportunity to execution with greater speed and confidence than any manual workflow ever could.
Beyond the hackathon, the goal is a product that users pay for because it delivers measurable value — a sustainable business capable of generating revenue on its own terms.
What started as a personal challenge to earn $$1000$ is evolving into a platform designed to help creators turn every opportunity into an outcome.
Built With
- bountypilot-ai-was-built-using-qwen-cloud-for-reasoning-and-workflow-intelligence
- dextopus-for-payment-structure
- express.js-for-backend-services
- github-for-version-control
- namecheap
- novus
- novusai
- playwright-and-beautifulsoup-for-opportunity-extraction
- postgresql-with-drizzle-orm-for-data-persistence
- react-and-typescript-for-the-frontend
- replit-for-development
Log in or sign up for Devpost to join the conversation.