Tian2 builds with artificial intelligence the way it builds everything else — as craft kept to a standard. A library of original coaching skills, sorted here into clear categories; a tool that turns a solution book into an online question bank for students; a curriculum that teaches the field from first principles; and the mentored research it all serves. Tools and teaching are Tian2's own work; student research is shown by field alone, names withheld.
The skills, the curriculum, and the animation work are Tian2's own — original, publishable, and credited to the studio. These are shown by name.
Student research belongs to the students. It appears here by field and method only — no names, no scores, no project identifiers. When in doubt, it is left out.
The clearest example of AI put to work for a student's own learning: a skill that takes a finished LaTeX solution book and turns it into an interactive, offline training question bank — the same source that sets the printed Physics-Challenge volumes becomes something a student can drill in the browser.
A printed solution book is a fine reference and a poor drill. This skill reads the book's LaTeX source — every question, answer, and worked solution — and builds an interactive online question bank from it: searchable, filterable by topic and difficulty, with each answer held behind a reveal so the student attempts the problem first. It runs offline from a single folder, and it keeps the mathematics intact, which is the whole engineering problem — fractions, exponents, and every alignment survive the conversion instead of collapsing into raw code.
This is how Tian2's Physics-Challenge books become training tools: the same corrected solutions that go to print also become the interactive bank a student practises against.
Parse each question / answer / solution block from the LaTeX source, protecting every equation before a single transform runs.
Emit a searchable app, a browse site, and a Markdown bank — filterable by topic and difficulty, answers behind a reveal.
The whole thing, MathJax included, runs from a folder with no server — a student practises anywhere, math rendered correctly.
A working library of more than forty-five original Claude skills — small, reusable agents that run a bilingual coaching and publishing practice. Rather than list them in one heap, they are sorted here into five categories by the job they do: competition and science-fair coaching, S.T. Yau award coaching, book writing and typesetting, teaching and curriculum, and applied AI. A representative skill in each category has its own worked demo.
The ISEF and science-fair pathway, packaged as tools. These skills take a student from choosing a topic and clearing the compliance paperwork, through data analysis and poster design, to a full judge simulation and interview drilling — each grounded in the official rules and rubrics, never invented. Published as an open-source plugin: ISEF Research Suite (MIT, 12 skills) · Install & details →
Reads a poster or slide deck, runs parallel agents to check the science against current literature, scores it against the official rubric, and returns categorized judge questions with strong and weak model answers.
View demo →Walks a project through ISEF's compliance intake — humans, animals, tissue, hazards, field work — and emits the exact forms, the correct review body, and the right pre- vs post-experimentation timing.
View demo →Discovers or scores a research topic for ISEF and high-school fairs, cross-validated against OpenAlex and domain databases, weighting conceptual creativity and surfacing compliance flags inline.
Drafts the ISEF Research Plan in the exact SRC pre-approval format, prompting each of the seven sections to extract content from the student — then running a critique pass for missing rigor.
Right-sizes a student's statistics so they pass judging without over-engineering: infers the design, recommends the correct test, and writes a methods paragraph with proper error bars and effect sizes.
Produces a printable, editable ISEF poster fitted to the official booth dimensions — LaTeX or PowerPoint — validated against the Display & Safety rules at print-ready 300dpi.
Critiques and tightens an ISEF abstract against the official 250-word, one-page constraint and the poster rubric — flagging methods-bloat, vague significance, and missing impact.
Helps a student find a research mentor by topic via OpenAlex and drafts non-spammy outreach locally — never auto-sending, never scraping private addresses, with strong ethical guards.
Maps the affiliated-fair path from a student's region to ISEF qualification — CASTIC and the 英才计划 for China-track students, regional → state for US students — with a personalized timeline.
Briefs a student on what their ISEF category's judges typically care about — grounded in the official Grand Award criteria — plus tactical guidance for the 15-minute booth interview.
Generates a bilingual 60+ question interview-prep document across every judge category, with strong model answers in English and Chinese and explicit "answers to avoid".
The feedback loop: consumes the student's annotated prep document and responds question by question — factual fixes, phrasing improvements, and clear yes/no judgments.
A full nine-skill suite for the S.T. Yau High School Science Award — a harder, research-paper competition judged by international scientists at an all-English defense. The skills carry a student from choosing a subject and a topic, through the paper and its data, to mentor outreach and the finals defense — grounded in a local archive of 393 past winners and the award whitepaper, and built around Yau's 自己出题自己做 ethic. Published as an open-source plugin: Yau Research Suite (MIT, 9 skills) · Install & details →
Discovers or scores a Yau research topic across the six subjects, weighting conceptual originality and verifiable student ownership above data scale — grounded in the winners archive.
View demo →Runs a fit interview to help a student choose among the six Yau subjects, briefing each panel's temperament and winning patterns — including the cross-discipline 科学金奖.
Structures the paper outline to the chosen subject — a math proof, an experiment, a CS system, an econ model each need a different skeleton — with an originality-and-feasibility critique.
Guides the actual paper to Yau norms — structure, English quality, citation honesty, a concrete abstract — coaching the student to write in their own voice for the finals defense.
Right-sizes the analysis per subject and flags the landmines judges hammer — correlation-as-causation, no validation set, weak baselines — without over- or under-engineering.
Prepares a finalist for the all-English oral defense at the Tsinghua finals: PPT structure, per-subject question drills, whiteboard-derivation readiness, and English-presentation tactics.
Guides a student through the award's implicit AI-use norms — when and how to disclose AI in the paper, the per-subject etiquette, and the judge Q&A landmines — framed as inferred, not stated.
Helps find and ethically approach a mentor while enforcing Yau's ownership ethic — the student leads, the mentor is auxiliary — with local-only outreach drafts and no email scraping.
Explains the Yau pathway end-to-end across the three regional tracks — 报名 through 分赛区, 半决赛, and the all-English 总决赛 — grounded in the whitepaper, never inventing a deadline.
The pipeline behind the printed solution books and their online question banks. These skills take a competition's past papers all the way to a typeset, categorised volume: extracting questions, generating and cleaning solutions, restructuring raw Mathpix LaTeX, building the book, verifying the math against source, and — the flagship — turning the finished book into an interactive online question bank.
Turns a LaTeX solution book into a filterable, offline, reveal-on-click online question bank — protecting every math form so fractions and align environments survive the conversion. The AI-for-Learning feature above.
View demo →Turns any exam's past papers into a priority-tiered "how to score more" guide — key facts per theme, the most common trap, a year-by-year question map — and optionally a corrected prep book.
View demo →The end-to-end workflow for a competition solution manual — orchestrates extraction, solution generation, integration, cleanup, and compilation across a whole year of questions.
Generates detailed step-by-step LaTeX solutions for olympiad and contest questions — multiple choice and long-form calculation problems with parts, ready to drop into a book.
Extracts questions from competition LaTeX into a structured format for solution generation or analysis — preserving images, options, and formatting across multiple question types.
Restructures and renames raw LaTeX from PDF extraction (Mathpix output) — flattening UUID folders and renaming files and images by problem section, e.g. "A1", "B2".
Cleans and fixes LaTeX before compilation — removing boxed answers, closing unclosed environments, and clearing placeholder answer blocks that break the build.
Converts restructured LaTeX into the book template's question/solution environments — handling image paths, display math, and template integration into a compilable book.
Builds or extends a past-papers olympiad book: first-principles background chapters, then year-by-year papers with Q/A/Solution blocks, cross-references, and the typography traps that waste hours.
Verifies and fixes USAMO problem and solution LaTeX against the AoPS wiki source — cross-checking AI-generated content against the authoritative original.
A personal fork of any2pdf that adds H4–H6 heading support — typesetting long Chinese books with deep heading hierarchies to PDF where the upstream tool hangs.
The skills that turn material into something teachable and shippable — interactive lecture decks in the house design system, EPQ document formatting to AQA standard, mindmaps and class summaries, the brand system itself, and the finalize-and-review pipeline that readies a deliverable for release.
Turns notes or an outline into a self-contained interactive slide deck — in-class handwriting annotation, one-slide-per-page PDF export, live canvas diagrams — running by double-click, no server.
View demo →Applies the Tian2 brand — the warm cream/ink editorial palette, New-Yorker-influenced typography, retro flat 2.5px illustrations — to any artifact: slides, web, print, or a book.
Formats EPQ Word documents to official AQA/A-level standards — Times New Roman 12pt, justified body, left-aligned subsection titles, the correct first-line indent — filling templates from a draft.
Generates a structured EPQ mindmap in Markdown — supporting both the academic-research and experimental-STEM formats — to scaffold a project before the writing starts.
Condenses a class or lesson into a clean structured summary — the notes a student keeps after the session, kept in the house style.
The pre-launch finalize-and-review pipeline for a deliverable tree — improves PDF appearance, refreshes derived assets, then runs a four-lens review and fixes the findings.
Evaluates scholarly work with the ScholarEval framework — structured assessment across problem formulation, methodology, analysis, and writing, with quantitative scoring and actionable feedback.
The rest of the working kit — modeling-competition model selection, a Chinese humanizer that rewrites in the studio's own voice, AI image generation for book covers, and the store-and-listing tooling that packages a finished deliverable. Applied AI, kept to the same standard as the rest.
Parses a modeling-competition problem PDF or Word doc, extracts the task and data conditions, and recommends which model or method to use, with a validation route.
Generates a modeling-paper structure and writing checklist against common scoring points, and recommends model selection and a control-experiment route by problem type.
Rewrites AI-drafted Chinese into the studio's own literary voice — first stripping the generic AI register, then layering the house writing principles and forbidden-phrase list on top.
Generates a professional book cover with AI image generation — combining a subject photo, a painterly style reference, and a template layout into a complete jacket.
Turns a set of PDFs into a complete, copy-paste-ready Chinese marketplace listing — title, attributes, preview images, description, and a price table — for RedNote and similar.
A workflow utility for the counseling vault — creates a new student workspace by duplicating the template structure and populating it from input files. Studio operations, not student data.
Three things built with and about AI. The first is a full ten-lesson curriculum that takes a high-school student from "what is a function" to a deployed speech system; the second uses AI-assisted animation to make physics move on the page; the third is a reference — a searchable dictionary of the Chinese that developers speak. The curriculum design, the animation code, and the dictionary are Tian2's own; the student project the course is built around is shown only in outline, its author withheld.
A bilingual course Tian2 designed and guided, built around a real student project: a Ningbo-dialect (Wu / 吴语) text-to-speech system for elder care. Rather than teach machine learning in the abstract, each lesson advances the same project — duration models, embeddings, attention, the physics of sound, modern neural speech, the ethics of data, and a responsible launch. It ships as ten lesson plans, ten lecture scripts, and ten interactive decks in the house design system — built with tian2-html-interactive.
A growing codebase that pairs AI with Manim — the programmatic animation library — to build animated physics lectures from scratch. The first unit, on kinematics, is complete: motion, velocity, and acceleration drawn as they actually behave, with lecture notes written alongside the code. It is early work, and openly so — a study in using AI to produce teaching material that holds to the same clean, unhurried standard as the printed books.
The last lesson of the curriculum is the one Tian2 cares about most: consent for elder subjects, audible AI-speech disclosure, audio watermarking, and the voice-spoofing law across three jurisdictions — because shipping responsibly is part of the syllabus, not an afterthought.
A linguistic dictionary of the Chinese developers actually speak — 变量, 句柄, 钩子, 幂等. Two hundred thirty-five entries across eight chapters, each read for its literal characters, its technical meaning, a real usage example, and the word-formation logic — calque, coinage, borrowing — that produced it. Written with AI as a lexicographer's assistant and set in its own searchable, day-and-night edition: a study in how a field quietly builds its own language.
The projects below belong to the students who did them. They are described by field and method only — no names, no scores, no project identifiers, no review content. This is a sample of the kinds of AI-assisted research Tian2 mentors, not a roster.
Real-time video segmentation and document-image (OCR) pipelines — deep models built, trained, and evaluated against published baselines.
CVPR · Science FairClassification and risk-modeling projects on real scientific data — exoplanet detection from light curves, clinical risk estimation — with honest treatment of imbalance and uncertainty.
ISEF · Science FairBioinformatics and machine learning applied to cancer immunotherapy — sequence and expression analysis read for signal, framed for a research review.
ISEF · YauAgent-based and learned simulation — pedestrian evacuation dynamics — alongside dialect speech synthesis, where the modeling and the ethics are coached together.
ISEFThese archetypes are part of a wider mentoring practice across the sciences. See the full Research Mentoring section →
Not Tian2's work — open systems kept close and read for what they get right. Each one sits next to a part of the practice: autonomous research next to student mentoring, poster and slide generation next to competition coaching.
A fully autonomous research agent — hypothesis, experiment, analysis, manuscript — whose work first cleared peer review at an ICLR 2025 workshop. Studied as a mirror for how student research is structured and defended.
A multimodal system that generates a conference poster directly from a paper. Read closely for the same reason the studio coaches posters by hand: the hard part is deciding what to leave out.
A multilingual agent that drafts presentation slides across twenty-plus languages. A useful counterpoint to the studio's own design-first approach to teaching decks.
AI can draft a thing quickly. Keeping it correct, and honest, is still the work.
— The Tian2 Editors