Custom Cultivation Program

Not more class hours —
a cultivation system built around one learner

Integration Track

A custom cultivation system with the growth archive as its spine: theory · math · practice · project, running alongside (not instead of) your current program.

Step one is always the precision diagnostic

Fully credited toward the program. It turns "unmeasured" into "named". If it ends here, the data is still yours.

Four layers

Theory11 modules

Published, bilingual system texts; every chapter requires a self-explanation artifact

  • Mathematical Thinking · Problem Understanding & Modeling60m
  • Mathematical Thinking · Logical Reasoning & Proof60m
  • Mathematical Thinking · Pattern Recognition & Abstraction60m
  • +8
Math11 modules

"One math problem → one AI project" five-segment cards + diagnostic-driven interleaving

  • Precision Diagnostic (four domains + proof + time profile)90m
  • Five-Segment Card · Sequences → PageRank (open sample)45m
  • Five-Segment Card · Derivative → Gradient Descent45m
  • +8
Practice5 modules

From knowing math to building agents: data literacy and engineering scaffolds

  • How to Ask AI Without Losing Your Skills45m
  • Python Data Literacy (toolchain → EDA)60m
  • Agent Engineering Starter (hands-on scaffold)60m
  • +2
Project3 modules

Graded project pool → original capstone → versioned portfolio

  • First L3 Real Project90m
  • Semester Mini-Capstone → Portfolio v190m
  • Hatch · Train a Net That Reads Digits120m

The AI protocol — where we differ from most AI courses

  • Solve before tool: core math is hand-worked (no AI) and archived first; the AI-assisted version comes after — the diff is verifiable learning evidence
  • All formative assessments are no-AI (RCT evidence: unguarded AI tutoring erodes independent skill)
  • The first lesson is not a tool — it is how to ask AI without losing your skills
  • Every project ships with an AI-use declaration in the archive

What we measure, and how

  • Completed ≠ passed: passing requires a delayed retention check (2-4 weeks) plus a transfer item
  • Cross-domain formative assessment every 4-6 weeks; semester re-diagnostic with a versioned ability profile
  • The parent monthly report is human-verified: retention and profile deltas, not completion percentages

Weekly commitment

About 4.5-5.5 hours/week in term (hard cap 6), scheduled explicitly around school and your current program. Under pressure we cut new material first and protect review; exam weeks, this program yields.

Division of labor with your current program

Advancing the curriculumYour programWe do not duplicate
Integration (proof · cross-topic · multi-representation)First casualty of large-group pacingOur mainline
AI trajectory (math → algorithms → projects → portfolio)Outside its structureOur destination
Growth archive (ability profile · versioned work · verifiable record)A report cardOur spine

Choose your depth

Recommended
Core Program · annualMember pricing
  • All four layers unlocked (theory · math · practice · project)
  • Weekly orchestration reviewed and signed by a learning designer (48h response)
  • Monthly 30-min 1:1 planning call (12/yr) + 2 human capstone reviews
  • Parent monthly report, human-verified: retention & profile deltas, not completion %
  • Assessment cadence: no-AI cross-domain check every 4-6 weeks + semester re-diagnostic
  • Growth archive: points · badges · versioned portfolio
By application · limited seats
Mentor Program · annual (by application)Member pricing
  • Everything in Core
  • Weekly 60-min 1:1 mentorship ×40 + weekly async work review
  • 8h capstone intensive + 4h application-season portfolio curation
  • Competition module included (AMC→AIME ladder as an external check of integration)
  • Demo Day showcase (annual student exhibition)
  • Capacity: ≤6 students per mentor

The mentor tier is application-only; submit interest and we schedule a conversation.

寻找全程私教的旗舰路线? 旗舰私教线 Founder Track →

Not sure? Start with the Starter Blueprint: a plan plus 4 weeks actually executed, fully credited toward an annual tier within 90 days. ask your advisor

The System · Math → AI → ML / Neural Nets

Five layers, each gated by mastery evidence — never by attendance. Fifth generation of the system, evolved from the first book series.

  1. ① Math foundation

    Mathematical thinking · proof line · interleaved practice

    From competition inequalities to convergence arguments, all archived

  2. ② Bridge cards

    Sequences→PageRank · derivative→gradient descent · matrix→forward pass · chain rule→backprop · probability→Bayes

    Every card is solve-before-tool; the hand-worked page archives first

  3. ③ ML / neural-net track

    Loss & generalization → backprop theory → why architectures exist

    A from-scratch NumPy net, then a real training loop dissected line by line

  4. ④ Hatch project (talent-routed)

    Digit-reading net / the music line: Fourier → note recognition → a net that hears instruments

    Hand-worked version first; the AI-assisted one is a NEW version — the diff is the evidence. Talent picks the track, not just the résumé

  5. ⑤ Portfolio → competitions & camps

    AMC/AIME ladder · proof-camp and research-program applications

    Targets every October, applications in winter; the portfolio IS the evidence behind every recommendation

Take the Challenge Set · find your entry point

Competition-band goes straight to the ML track; anyone else learns exactly which three cards come first.

What you buy is what your child builds

Project-based learning is the core deliverable: at least one real project per semester, hatched into the versioned portfolio. The full catalog is public, maturity honestly labeled.

Browse every project →

Commitments & terms

The guardian account manages and pays; the student gets their own learning view and portfolio. Payments are hosted by Stripe and never touch this site.

Every pedagogical claim on this page is evidence-graded (mastery learning, deliberate practice, spaced retrieval, interleaving are all high-evidence); the portfolio's admissions value is a reasonable inference — we make no individual-outcome promises.

program: pathfinder