CJ/HERESEARCH INTO REALITY中文
CJ/HEResearch · Mathematics · ProductsSelected work

CHANGJIANHE.

Changjian He
AI RESEARCHER
APPLIED MATHEMATICIAN
FOUNDER

FROM IDEA TO IMPLEMENTATION

Intelligence,
built for the
real world.

Explore all projects
CJ/HE01 / Applied intelligenceProduct chapter
New York & New Jersey

A city of options.
A clearer choice.

A rental-listing agent that brings inventory, location, and commute context into one workflow.

INTERACTIVE SAMPLE · FICTIONAL LISTING
NYC / NJ · RENTAL EXPLORERJERSEY CITYMANHATTANBROOKLYN
Sample 01 / Jersey City

$3,200 / mo

1 bed · In-unit laundry
Commute context available

Schematic geography · Illustrative route
Discover →︎ Compare →︎ ShortlistExplore the rental demo
CJ/HE02 / ElementizationConceptual visual

Change the substrate.
Define what’s possible.

Structured data becomes computational elements for a declared scope of operations.

ELEMENTIZATIONPrivate transformationSTRUCTURED RECORDSINTERNAL METHOD OMITTEDCOMPUTATIONAL ELEMENTS
Structured recordsInside an organization boundary
ElementizationPrivate transformation · Internals omitted
Computational elementsFor a declared scope of operations

Compare elements using the declared geometry. Visual positions are illustrative.

Explore Elementization
CJ/HE03 / Elementization StudioSanitized walkthrough
From theory to a working tool

Inspect.
Evaluate.
Then approve.

A local desktop workflow with explicit review, validation, and human approval.

MVP-A · Implemented desktop workflow
Release checks remain at the recorded checkpoint.

INNERFY / STUDIOIllustrative workspace
LOCAL WORKSPACE

Begin with structure.

Inspect a structured dataset before choosing how to proceed.

Interface interpretation based on implemented screens. Sample states, not a live certification run.

Local processing · Human review · Governed artifactsInside the Studio
CJ/HE04 / Machine learning securityResearch chapter
ALOA / Two-tower models

What does
a model
remember?

Inferring training membership from embedding-based recommendation models.

User and item towers turn their respective inputs into embeddings.

USER INPUTITEM INPUTCONTINUOUS EMBEDDINGSUser feature changed
Reported shadow-based evaluation / Table 6
Combined
97.87%
Dummy
95.07%

Accuracy · Different class balances: 4:1 vs 1:10 IN:OUT.
These are reported experimental results, not universal success rates.

Open the full research paper
Conceptual geometry · Not measured embedding coordinates
Full story: towers →︎ shadow model →︎ perturbations →︎ classifier →︎ evidenceExplore the research
CJ/HEIdeas deserve implementation.Closing chapter

Let’s build
what’s next.

Applied AI, research collaborations, and ambitious products that need technical depth.

Let’s connect