AI applications

AI features backed by shipped systems — and an honest list of what we don't do.

Our AI work concentrates on two lines with shipping evidence: wrapping large-model capabilities into unified REST services that production systems call, and phone-based 3D scanning apps live on the US App Store. Everything we list here has been deployed; everything that has not — model training, speech recognition — we say so up front.

01

What we build

LLM integration & serving

A large model (GPT-4o vision, Qwen) wrapped into one REST service with timeouts, retries and fallbacks — optionally behind an OpenAI-compatible interface — and reused by every client: web, mini program, PC and admin.

3D scanning and reconstruction apps

Native iOS apps that drive the depth camera, reconstruct a measurable 3D model and upload it for further processing. Two such apps are live on the US App Store and still updated.

AI data Q&A (POC)

Business users ask in plain language; the system generates read-only SQL over whitelisted tables and shows the SQL for review. Runs on a local model inside your intranet. Implemented in our own demo environment; production rollout is assessed per client.

Feasibility first

A small POC on your own data with an agreed evaluation set, before anyone commits to a full build.

02

See it run

Internal demo implemented on our own ERP product: ask in plain Chinese, a locally running model generates read-only SQL, the result comes back with the SQL shown for review. Fictional data; deployable inside your intranet with the model running locally.

02

How we keep AI projects honest

Data stays inside your intranet

Models can be deployed locally (vLLM serving open-weight models such as Qwen). What may leave the building is decided in the design document, not at runtime.

A human in the loop

Where an error is expensive, the model assists and a person confirms. The review step is part of the delivery, not an option.

Measured on an evaluation set

We report results on a test set agreed in advance, not on a demo impression — and we do not promise accuracy numbers.

We name what we don't do

Model training and speech recognition are not on our list: no shipped evidence, no offering.

03 — Relevant work

Where this shows up.

Fish recognition API: a vision LLM wrapped into one recognition service
AI & vision

Fish recognition API: a vision LLM wrapped into one recognition service

An aquaculture technology company

A REST service that turns fish photos into a normalized species name: a vision LLM analyses the photo and the Go back end extracts the standard name. Deployed to production and reused by a web front end, a mini program, a PC site and an admin console through one API document.

Go · GPT-4o vision · REST

Full write-up (Chinese)
Foot-scanning apps for a US foot-care company
Native appsAI & visionApp publishing

Foot-scanning apps for a US foot-care company

A US foot-care and orthotics company

Two native Swift apps that capture the foot in 3D and turn the scan into a custom orthotics order. Released on the US App Store, with the first app maintained and updated since 2022.

Swift · 3D scanning SDK · Firebase

Full write-up (Chinese)
3D scanning and reconstruction pipeline
AI & visionNative apps

3D scanning and reconstruction pipeline

A 3D scanning technology company

An end-to-end chain from phone capture to finished 3D model: capture happens on the device in Swift while heavy reconstruction runs server-side in C++, followed by Python post-processing — textures, mesh reduction, format conversion — and a Java + Vue platform that manages and distributes the models. Eight repositories, maintained from 2024 to 2026.

Swift · C++ · Python

Full write-up (Chinese)
Interactive social video app for iOS
Native appsAI & visionApp publishing

Interactive social video app for iOS

An overseas social video start-up

A native Swift app, from version 1.0 in 2024 through a full 2.0 rebuild, together with a Python model service for image-to-text, video and music generation. Released on the US and UK App Stores in 2026 and still being updated.

Swift · Firebase · Python

AI chat and content-creation apps
AI & vision

AI chat and content-creation apps

An AI application company

Three front-ends for AI products (conversational, form-driven and creative) on web and cross-platform mobile, maintained from 2024 to 2026.

Vue · uni-app · LLM APIs

Full write-up (Chinese)
Photo-based species identification app
AI & visionNative apps

Photo-based species identification app

Our own product

Take a photo, identify the aquarium species, and land on its encyclopedia entry and community posts. Recognition, content and admin share one data model.

Image recognition · Mobile

Full write-up (Chinese)
04 — FAQ

Questions we get asked.

Will our data leak to third parties?

It does not have to. Models can be deployed locally inside your intranet — our AI data-Q&A demo runs on a local model with a read-only SQL whitelist and no external calls. When a third-party API is the better choice, which fields may be sent is fixed in the design document before development starts.

How fast can you prove it works?

Proof-of-concept pace: pick one workflow, build a small evaluation set from your own data, run it, and judge from the measured results. A POC is measured in weeks and does not commit you to a larger contract.

Do you train models?

No. Our work is integration, service packaging and the engineering around existing models. We do not offer model training or data labelling, and we say so up front rather than take the project and improvise.

What shipped AI work can you show?

Two lines with production evidence: a vision-LLM recognition service deployed in production and reused by several client front ends, and phone-based 3D scanning apps live on the US App Store. Our AI data-Q&A demo (local model, data stays on-premises) runs in our own environment.

Can you guarantee recognition accuracy?

No — and we put that in writing. We deliver the evaluation set and the measured report instead, and keep human review where mistakes are costly, so a wrong model output cannot reach production data unconfirmed.

Do you do speech recognition?

We have no shippable evidence for ASR or TTS, so we do not offer it. Real-time voice chat and live-streaming products are a different line we have shipped — see Native apps.

Contact

Tell us what you want to build.

What the product does, which platforms or devices are involved, and your timing. We reply within three working days, in English, with questions or a short written proposal.

Based inBeijing, China (UTC+8)