ai.op-log.gg

Inspiration Collectible card markets move like stock markets, with spikes, crashes, and hype cycles, but collectors trade them blind. A One Piece TCG card can double in a week because of a tournament result or an anime episode, and the only "research" available is scrolling marketplaces and guessing. Stock traders get Bloomberg. Card collectors get vibes.

We'd already built the data side: op-log.gg, a market-analytics site tracking every One Piece card daily. The XPRIZE question was the leap. What if the analyst, the one who reads all that data and tells you what it means, was an AI? ai.op-log.gg is that analyst, built as a new product on top of the existing data engine (reuse disclosed in the repository's initial commit).

What it does Ask it what a stock analyst gets asked: "Is OP08-098 undervalued?" "What should I sell from my collection?" "What's driving the market this week?"

Gemini answers with function-calling over 13 live read-only tools: current prices across 6,671 cards, an ML fair-value model, community sentiment, graded population counts, live marketplace listings, tournament meta usage, and (for signed-in users) their own collection and watchlist. A strict tool-or-silence rule governs every reply. Any factual price or value claim must come from a tool call, and the model is not allowed to remember, estimate, or invent a number. Answers stream token by token, every card mentioned becomes a link with live purchase options, and chat history follows the account across devices.

Revenue is real: in a single recent week our affiliate links drove over $10,000 in marketplace card sales at an 8.5% click-to-purchase conversion rate, generating a few hundred dollars in commission payouts, plus early commissions from a second retail partner. The subscription funnel is live end to end and produced its first paid conversion (later refunded), validating that the upgrade path works.

The AI doesn't just answer questions. It runs the operation. The chat is the visible AI. The business underneath is operated by AI systems that make decisions daily without a human in the loop:

A local LLM scores community sentiment around the clock across 16 card games, working through thousands of Reddit, YouTube, and social items daily. Scores are blended with a deterministic keyword anchor so a hallucinated score can't move the published index. A gradient-boosted fair-value model (74 features covering prices, supply, sales velocity, graded populations, and character popularity) retrains weekly and prices the catalog nightly, with conformally calibrated confidence intervals. Low-confidence predictions suppress themselves. The model grades its own homework. Every prediction is scored against the realized market price seven days later, and the results, including the misses, are published daily on a public /model-performance page. We believe we're the only product in this niche that publishes its model's error rate. Automated pipelines watch marketplaces for mispriced listings, ingest tournament decklists, and monitor their own health. Every job writes an audit trail, and a watchdog messages the operator only when something breaks. Human vs. AI responsibilities, honestly drawn: the AI executes pricing, scoring, predicting, self-evaluation, and customer Q&A every day. The human makes product decisions, resolves judgment calls the data can't settle (a small override map exists precisely for cards where two data sources disagree and only a person can adjudicate), and approves model retrains before they change every number on the site. The AI operates; the human governs.

There's a second layer: this product was largely built by AI coding agents directed by a solo founder. The same AI-native principle applied to construction, with every change audited, tested, and reviewed before shipping.

Challenges Gemini's thought signatures. Our first tool-calling loop hand-built the conversation history, which silently broke Gemini 3.x's reasoning-signature chain. The model would call tools forever and return empty answers. The fix (use the SDK's chat session so signatures survive) is invisible in the product and was brutal to diagnose.

Thinking vs. answering. Uncapped, Gemini spent its entire output budget on internal reasoning and returned empty text. Capping the thinking budget guaranteed room for the reply.

Making a hard quota fair. Pro AI's 100 messages anchor to each subscriber's own billing cycle rather than the calendar month, so someone subscribing on the 25th gets a full allowance instead of a five-day stub. The free trial is lifetime-tracked separately so subscribe-then-cancel can't farm fresh trials, and quota is reserved before each Gemini call so concurrent requests can't slip past the cap.

Trusting an unattended AI. The real work of AI-native operations is boring: heartbeats, audit documents, fail-closed guards, a model that refuses to publish when its inputs are incomplete. Autonomy came from making failure loud, not from making the AI smarter.

What we learned Grounding beats generation. Retrieval with a model forbidden to invent numbers earns trust that a fluent guesser never will. And publishing your error rate is a feature: collectors don't expect a crystal ball; they expect honesty about uncertainty.

Economic opportunity Beyond subscriptions, the platform runs a creator partnership program. Content creators get tracked referral links and a profit-share pool split by the traffic they drive, turning community YouTubers into stakeholders in an AI-run business.

What's next Deeper attention signals feeding the model, richer portfolio advice, and, after the competition, bringing the AI analyst to the flagship site's full audience.