Beidashi’s AI Decision-Making Tool: Cold Start and Data Governance Experience

AI Summary · Serial Entrepreneur Perspective（Summarized by AI; views belong to the original author. You can skip the full article after reading this.）

Developed by a Peking University PhD candidate addressing his own pain points, AXIA builds a mentor database of 7,265 advisors via crowdsourcing to solve information asymmetry. Its core strengths lie in data governance—refusing automatic merging and distinguishing citations from inferences—and clear boundaries for AI usage (CoT is disabled for pure retrieval and structured tasks to cut costs and speed up responses). The opportunity is real and suits teams with academic resources or access to vertical communities. The biggest pitfall is weak willingness among individual users to pay, plus the difficulty of standardizing unstructured information like character traits and supervision intensity.

• Data governance: refuse to merge names based on pinyin alone, isolate by default…
• Cost reduction through tech: disable model reasoning for structured tasks (retrieval, extraction, summarization), cutting inference time from 40 seconds to 5 seconds.
• Crowdsourced cold start: automatically expand the database using user search behavior (create a record the moment a new university is looked up) …
• Information tiering: clearly separate “officially archived” data from “inferred from paper authorship”…
• Business model reference: basic searches are free to drive traffic; premium zones (senior Q&A / writing workshop) use managed paid services to build a trust loop.

1. What kind of opportunity is this?

AXIA, developed by a PKU PhD candidate to solve his own pain points, is an “AI advisor matching” tool tailored for graduate students. It builds a mentor database through crowdsourcing—having recorded 7,265 advisors and 20,995 papers as of September 2026—to tackle the confusion and information asymmetry surrounding supervisors during the admissions process. Its business model follows a “freemium + premium” strategy: basic lookups are free, while deeper services like senior student Q&A and document workshops charge fees, creating a trust loop.

2. Independent judgment

Worth pursuing, but the barrier lies in governing vertical data, not in acquiring traffic. This direction targets the extremely high “mismatch rate” pain point in academia, where existing public data is highly chaotic. The core moat is not the algorithm, but the strict data governance standards the author has built—such as refusing automatic merges and distinguishing inference from fact. Inference: Tools like this naturally suit teams with university connections or access to vertical communities. Pure tech teams, lacking an understanding of “unwritten academic rules,” will very likely fall into the quagmire of messy, unreliable data.

3. Cold-start path

First validation step: Automatically expand the database through user search behavior. When a user looks up an uncovered university, the system creates the record, scores it, and stores the evidence right then and there. The next searcher will immediately see the result. This “build-as-you-look-up” model achieves zero-cost coverage of long-tail universities.
Cost scale: Minimal; the primary expenses are server hosting and model API calls.
Timeline: Depends on seed users actively contributing. In the early stage, the developer or a small team can validate the core database framework within 3–6 months.

4. Biggest risks and how to avoid pitfalls

1. Weak willingness among individual users to pay, plus compliance risks: Students are price-sensitive, and the data involves privacy. How to avoid the trap: isolate unstructured information (character traits, supervision intensity) from public scores, turning it into paid anonymous Q&A managed by the platform. Let the platform handle payments to build trust, and avoid selling sensitive data directly.
2. Collapse of trust caused by inaccurate data: The academic circle has many name collisions and confusing identity markers. How to avoid the trap: strictly enforce the “refuse to merge” principle. Default to isolating users with the same name unless a single personal homepage serves as ironclad proof. If in doubt, split into multiple records rather than forcing a merge.

5. Case study (how others did it)

• Product positioning: Don’t aim to “recommend the best advisor”; instead, focus on “making the basis for judgment transparent.” Every piece of information must be clickable, showing the source webpage, crawl date, and confidence level, so users can make their own judgments.
• Data governance (hard rules): Extracted claims must include the original citation. If the source webpage cannot be searched to confirm the statement—even if it differs by a single word—the entire data entry is invalidated. Clearly distinguish between “we checked and it’s genuinely absent” and “we haven’t checked yet,” to prevent user misreading.
• Scoring system design: Abandon a single overall score. Display eight signals in layers: separate the advisor’s own attributes (hard facts) from student-advisor fit (dynamic results). Mark “inferred values” (e.g., signs of accepting students) in different colors and do not mix them with hard facts in weighted calculations. Items with insufficient information should not receive a zero score; instead, display “limited public information” and renormalize the weights of the remaining items.
• Cleaning up duplicate names / identical identities: Discovered that unique identifiers like ORCID had a 32% misbinding rate (e.g., a professor’s profile accidentally linked to another institute’s ID). Subsequent rules were changed: unverified IDs carry zero weight, and during verification, evidence brought by that very ID must be excluded to prevent errors from self-reinforcing.
• Practical cost reduction through tech: Discovered that “slowness” came from the model overthinking. Disabled Chain-of-Thought (CoT) for structured tasks such as planning, extraction, and summarization. Inference time dropped from 40.1 seconds to 4.9 seconds, and invalid reasoning portions were heavily trimmed from output tokens.
• Information tiering: Found that 97% of “affiliated institutions” came solely from paper authorship, not from official websites. After revising the rule, only information from official websites / personal homepages could be used for assertions; paper-authorship data was downgraded to “inference” and used only for recall hints.
• Customer acquisition and monetization loop: User profiling (automatic PDF resume parsing) is done only once. Matching is free. For questions like “character traits” or “ease of graduation” that lack public data, guide users toward paid “senior Q&A” (anonymous, platform-managed) and “writing workshop” (customized cover letters combining the advisor’s actual publication history), forming a complete journey from free filtering to paid decision support.

Original text ·少数派 -- Matrix：Read the full article →