CuratoBase
Mini App Curator
About
The problem it solves The Problem: The Base ecosystem is exploding with new applications, smart contracts, and mini-apps daily. For end-users, navigating this immense volume to discover high-quality, safe, and genuinely organic applications is incredibly difficult amidst the noise and engagement farming. Conversely, for authentic builders, gaining initial visibility for early-stage projects often requires heavy marketing budgets rather than just shipping excellent code. The Solution: CuratoBase solves this by deploying an autonomous AI Curation Agent that continuously monitors social feeds (Farcaster/Warpcast, Twitter) for builder activity. For Users: It intercepts "shipping" signals, rigorously scores them using a multi-tiered heuristic engine (checking for live demos, open-source repos, and organic reward distributions), and automatically surfaces only the highest-quality projects to an ultra-premium, noise-free "Prestige Registry" dashboard. For Builders: It creates an objective discovery pipeline. Authentic developers are rewarded with instant ecosystem visibility based purely on their shipping momentum, on-chain proofs, and community contributions; bypassing traditional marketing gatekeepers. Use of AI tools and agents CuratoBase is fundamentally driven by an Autonomous Curation Engine. AI is not just an add-on; it is the core infrastructure of the platform. Autonomous Polling Loop: A recurring agentic cycle awakens and scans external APIs (Searchcaster/Neynar) for specific ecosystem triggers ("deployed on base", "shipped v1", "airdrop"). Semantic Extraction & Scoring: An LLM is employed to perform intelligent sentiment analysis and semantic extraction on raw social data. It evaluates the project against our custom rubrics: Authenticity, Effort, and Verifiable Proof. The "Curator Verdict": Rather than just outputting numbers, the AI Agent generates a qualitative, human-readable insight (e.g., "Active builder consistently shipping updates. Latest signal is a live Base App"). Generative Trust Reports: The agent automatically authors complete, immutable "Trust Reports" documenting its exact reasoning for trusting or rejecting a project. This ensures AI decisions are completely transparent to end-users on the dashboard.
How it's made
Challenges we ran into Real-time Signal Noise vs. Authentic Shipping: Extracting genuine product launches from social media spam or engagement farming was highly complex. We overcame this by building a custom heuristic engine that doesn't just look for keywords, but demands verifiable proof (e.g., valid URLs, Github repo links, or on-chain transaction hashes) and cross-references them with known, reputable developer identities (FIDs) before escalating a signal. AI Hallucination & Deterministic Outputs: Initially, the agent would occasionally invent project names or misinterpret generic tweets as app launches. We solved this by anchoring our LLM outputs to a strict JSON schema constraint and requiring the agent to cite specific evidence (e.g., demo=YES, repo=NO) before a status could be upgraded to "CURATED". Dynamic UX without Hydration Errors: Passing asynchronous, real-time agent data to a premium frontend without causing UI lag or Next.js hydration mismatches was challenging. We resolved this by migrating entirely to Server-Side API routes connected to a Neon Postgres database, ensuring the UI always beautifully and immutably reflects the true state of the agent's findings. Intelligent Project Name Extraction: Raw signals often lacked clean formatting. We wrote an intelligent regex and context-parsing algorithm that reads the AI's natural language descriptions to dynamically extract real brand names (e.g., parsing "DropCast" from a messy tweet) rather than falling back to ugly database hashes