AI works because mathematics does.

The fine line between hallucination and prediction is not magic or a data dump. It is maths.

I engineer AI where the mathematics is explicit and the physics is enforced.

A metric that cannot see a fault will certify it. Those are the failures I build against.

0.000
change to a pose model’s original outputs after adding new keypoints Markerless 3D pose with public models: 7,000+
stores running the retail vision work Retail vision AI that scales without retraining: 1 of 3
global winners from 638 proposals TAITRA "Go Healthy with Taiwan" 2025 — the record

Two phone cameras capture a golf swing in 3D. Code measures it before a language model explains it.

Two ordinary phone cameras film a golf swing. A pose model finds points on the body and the club in each view, the two views are combined into one 3D skeleton at real-world scale, and a coaching system turns the measurements into advice. The language model in that system writes the advice; it never calculates a number.

IdeasLab Formosa, 2025 — present

Markerless 3D pose with public models

Adding new keypoints to a trained pose model while proving the old ones cannot change, and finding where 3D scores flatter the model.

Research engineer: experiments, training, evaluation

0.000
change on the original keypoints after adding new ones: 0.834 vs 0.733
AP on the new keypoints: side branch against a shared head

1 of 6The failure the validation score missedPassword-protected 2 of 6Adding keypoints without touching the old onesPassword-protected 3 of 6Reading the shape of the errorPassword-protected 4 of 6Labelling what the data does and doesn't knowPassword-protected 5 of 6Lifting to 3D, and what the public numbers sayPassword-protected 6 of 6Testing 3D models against independent measurementsPassword-protected

Fixing the camera geometry every number depends on

A 3D reconstruction without real-world scale gives sizes that are wrong but look right, which is how it gets past review.

Research engineer: finding faults, fixing them, checking the fixes at scale

4.8×
less reconstruction noise: 10–46×
more consistent bone lengths: −73%
error in event timing

1 of 4Bugs that produce believable outputPassword-protected 2 of 4Getting the scale rightPassword-protected 3 of 4Checking against real bodiesPassword-protected 4 of 4One threshold, two camerasPassword-protected

An assessment system where code calculates every number

Expert feedback from a language model, built so that it cannot make up a measurement.

Architect: rule engine, scoring, prompts, evaluation

0
numbers calculated by the language model: 8
languages, prompts included

1 of 4The core rulePassword-protected 2 of 4Same input, two pipelinesPassword-protected 3 of 4Testing the rulebook itselfPassword-protected 4 of 4Grounding and costPassword-protected

What others have published about this work

2025
XView AI, the swing-analysis app it shipped in
The pose and lifting research shipped as XView AI, a golf swing analysis app that needs no body markers and runs its whole analysis on the phone, with no cloud connection and no video leaving the device.
The 2D-to-3D pose lifting that the analysis takes its numbers from, and the CoreML deployment that keeps it all on the phone, are my work. The motion-capture case study above has the full account, including the trade-offs.
2025
Winner, “Go Healthy with Taiwan” 2025
One of three winners of TAITRA’s 2025 “Go Healthy with Taiwan” competition, picked from 638 proposals from 55 countries, for a baseball motion-analysis system built on XView’s motion-analysis technology.
I wrote the technical proposal, including the biomechanics requirements and the case for adapting the golf system to baseball. Market scope and standards came from work with the business team.
2026
Taiwan Expo Europe 2026, Warsaw
The motion-analysis work was shown under the Taiwan Excellence banner at Taiwan Expo in Europe, TAITRA’s main European event, on 22–24 June 2026 at EXPO XXI Warsaw. XView is named in the official exhibitor list.
I presented the technical case on the main stage to an international audience.
- Official exhibitor listing: Ideas Lab, Formosa, XView
- Taiwan Excellence: Expo Europe programme and dates

2025 —
Pose Lifting and Biomechanical Motion AnalysisApplied R&D at IdeasLab Formosa, in progress

hands follow the posegolf validation baselinebody pathway bit-identical

Two phone cameras, one body: points found in each view and combined into 3D, on the phone. Values here are qualitative. Drag to orbit.

The shelf-recognition system adds new products through its catalogue, without retraining either model.

Shelf recognition for more than 7,000 convenience stores. One model finds where each product sits on the shelf. A second turns each product into a vector and looks it up in a catalogue, so adding a new product means adding its vectors to the catalogue. Nothing is retrained.

President Information Corp, 2023 — 2025

Shelf recognition that learns new products without retraining

The catalogue adds new products by storing reference vectors, so neither model needs retraining.

Lead engineer: architecture, training, production deployment

7,000+
7-ELEVEN stores running it (Scientific Reports, 2025): +30%
recognition of products in a customer’s hand: Store #8
Taiwan’s 8th unmanned 7-Eleven, deployed

1 of 3Where it is, and what it isAn object detector finds products on the shelf.Read this section 2 of 3Designing for hidden productsIn a cashier-free store, products are recognised while they are in a customer's hand: partly covered, at any angle, and blurred by movement.Read this section 3 of 3Running it on NVIDIA hardwareProduction ran on NVIDIA's stack end to end, built together with NVIDIA: TAO Toolkit for training, DeepStream for video inference, TensorRT for fast execution on the hardware, and Metropolis microservices for cloud and edge deployment.Read this section

What others have published about this work

2023 — 2025
Running in more than 7,000 stores
President Information Corporation’s planogram-compliance system runs in more than 7,000 7-ELEVEN stores in Taiwan, according to its 2025 Scientific Reports paper.
I designed the detection and embedding architecture behind it: one model finds where a product is, a fine-tuned vector space says which product it is, and so a new product costs a few vectors instead of a training run.
- Scientific Reports (2025): real-time planogram compliance in 7,000+ stores

2025
Scalable Vision AI for Planogram ComplianceNVIDIA GTC 2025, technical poster, presented

products 14retraining runs 0off-plan facings 1

Left, the shelf a store camera sees; right, the catalogue of reference vectors. Each line is one lookup: the detector finds a product, and the nearest vector names it. The red can is named correctly, which is how the system knows it is on the wrong shelf. The two pairs of clusters that almost touch are flavour variants of one product. Drag to orbit.

Earlier work on attention, image generation and function approximation.

My thesis on filling in missing parts of images, a step-by-step mathematical walkthrough of the transformer, and an early experiment with Kolmogorov–Arnold networks. This is where I learned the tools the two projects above depend on.

NTUT and independent work, 2021 — present

Filling in missing parts of images, with little training data

With little data, a generator can collapse to a few outputs, memorise its training images, or never learn at all. All three give bad samples, and each has a different cause.

Author

1 of 2Two kinds of attentionContextual attention copies texture from related regions; spatial attention keeps structure aligned across the missing area.Read this section 2 of 2Why it still matters in my workThe attention studied here also appears in transformers.Read this section

2023
Content and Spatial Aware Generative Model for InpaintingMSc thesis, National Taipei University of Technology, published
2024
The Transformer Architecture — a mathematical walkthroughPublic technical writing, published
—
Efficient LLM Inference via Kolmogorov–Arnold NetworksIndependent research, in progress

rows queriescolumns keysupper triangle masked

In causal self-attention, each token can attend only to itself and earlier tokens. Denser cells mean more attention.

Lessons from things that went wrong

Each one came out of a failure in the work. They link to the case study it came from, the research note that develops it, or both.

Check that the test can see the failure

A pose model scored better while getting worse. Its validation set held only the narrow footage it was tuned on, where the new bias happened to be right.

From “The failure the validation score missed”Note: When the test set shares the blind spot

Rule out regressions by design

Freeze the original model's weights and its output cannot change, which a weight diff proves. A penalty term in the loss only makes a change less likely.

From “Adding keypoints without touching the old ones”Note: Adding an output without moving the old ones

Look at where the error is, not only how big it is

Error that grows toward the edge of what the network can see means it cannot see far enough; a bigger model will not help. Error that stays flat across sessions but groups by camera setup comes from how points were labelled, not from scale.

From “Reading the shape of the error”Note: What one cell of a network can see

Know which failures training cannot fix

A failure that only appears on footage unlike the training data is missing from the training and validation sets, so no loss function can see it. That fix belongs at inference.

Note: Losses and failures training never shows

Let code do the arithmetic

When a wrong number can hurt someone, Python calculates it and the language model only explains it.

From “The core rule”

Write down what is still broken

Every handover lists the failures that are still open, so the next person does not rediscover them the hard way.

Count independent sources, not samples

Samples from the same source share an effect, so a hundred thousand of them can be worth a few hundred. Put error bars on the number of sources.

Note: Effective sample size

Know what a guard does under your optimizer

A gradient clip that fires on every step fixes the step length under SGD and changes almost nothing under Adam. Under Adam its real job is the rare spike.

Note: Clipping under SGD and Adam

Score perfect inputs first

Run the evaluation on the true values before running it on a model. What remains is the pipeline's own floor, and no model can go below it.

Note: The floor of an evaluation

Compare scores as well as accuracy

The same weights gave logits seven apart on two execution paths while the accuracy matched. Thresholds and calibrated probabilities read the scores.

Note: Same weights, two answers

Illustration of the observed failure pattern. The curves show the relationship, not values from a logged run.

Things I build outside work

Hackathon entries and personal tools, each built to answer a specific question. Magic Shuffle also includes a deterministic fallback that works without a model or API key.

Biomech Emcee
2026, TypeScript
Hackathon entry
A browser workspace that a person and an AI agent operate together through WebMCP. The page exposes its live review state as 13 typed tools, nine that read and four that act (move the timeline, change what is shown, leave a note tied to a frame), so the agent works from the same view the person is looking at. A request the evidence cannot support gets a structured refusal instead of a plausible number. Biomechanics review is the reference workflow, and the analysis runs entirely in the browser.
WebMCP, Agent tool design, Structured refusals, React
CarbonPass
2026, Python
2026 Presidential Hackathon entry
Turns a Taiwanese factory’s photographed paperwork into carbon accounting: a product’s footprint on the EU CBAM template, where material is lost, production scheduling that follows the power grid, and anonymous comparison with similar factories. A vision-language model runs on the factory’s own hardware, so documents never leave the building.
Local VLM, OR-Tools MILP, FastAPI, CBAM
Magic Shuffle
2026, JavaScript
Hackathon entry
Picks a song from how you are doing (energy, sleep, stress), the weather and what is next on your calendar, and says why it fits. Spotify confirms that every track the model names exists, and a plain-code path gives results of the same quality with no API keys.
LLM grounding, Spotify API, Deterministic fallback
data_automation_pipeline
2025, Python
Personal tool
Converts PDFs, web pages and source files into clean Markdown to feed to language models, keeping LaTeX formulas and code blocks intact and tidying everything around them. Runs in parallel, with optional OCR for scientific PDFs.
LLM data prep, LaTeX preservation, OCR
CardPilot
2025, Python
Reads photos of business cards into structured records and stores them in a ChromaDB vector database, so you can search the collection in plain language instead of by exact match.
Gemini, ChromaDB, Semantic search, Gradio
Pulse
2026, Python · three.js
Hackathon entry
A shared 3D city on a projector at a live event. Guests scan a QR code and appear as characters; the expression on each face drives that person’s character and a mood panel for the room. The emotion is read on the phone and only its label is sent over the network. Built in a day.
Real-time, WebSocket, On-device inference, Privacy

From ERP software in Coimbatore to 3D vision in Taipei

2011 — 2014BSc, Information TechnologySri Ramakrishna Mission Vidyalaya College
2017 — 2021Software DeveloperAIBS Software Solutions
2021 — 2023MSc, Electrical Engineering & Computer ScienceNational Taipei University of Technology
2023 — 2025AI EngineerPresident Information Corp (統一資訊)
2025 —AI Research EngineerIdeasLab Formosa

What I did at each stop

What I’m open to

Research collaborations, hard applied problems in 3D vision or on-device AI, and conversations about either. I am based in Taiwan and open to relocating.