HappyDing - 技术博客

// SYS_STATUS: ACTIVE// LOCALE: EN-US

// ENGINE: NEXT_16// MODE: TECH_×_LIFE

Full-stack engineer, focused on AI

Read the blog About me GitHub↗

Building AI applications

0posts

0views

0categories

0tags

About me

I combine model capabilities, product sense and engineering systems to build AI applications that actually ship. Half of this site is engineering notes; the other half is life.

Next.js · React · TypeScript · Node.js · AI Agents

Education

Xinjiang Institute of Engineering

Data Science & Big Data Technology · BSc

2020.09 – 2024.06

Major focus: Machine learning · Deep learning · Big-data processing
Campus roles: Head of Youth League organization dept · Part-time deputy secretary

Experience

Ningbo Vorto Technology Co., Ltd.

2026.06 – Present

CTO

Set the technical direction and shipped the products: four lines built around applied AI — a model gateway, a foreign-trade prospecting SaaS, a multi-agent development framework and desktop tooling — from architecture choices and engineering standards through to containerised deployment.
Built an in-house AI token gateway so every model call goes through one place: forked and extended the open-source One API with group quotas, cost reconciliation, usage observability and alerting. Compliant models are served externally; internal projects onboard by being issued a group token — no gateway changes needed.
Took a one-stop foreign-trade prospecting platform from zero to production: FastAPI + PostgreSQL + Redis behind a React frontend, cross-platform images pushed to a private registry, containers orchestrated across a self-hosted NAS and a VPS reverse proxy, with automated daily database backups.

Projects

Vorto

CTO

liveAI × Product × People

An AI creative & software studio — AI creates with us, turning ideas into working products.

Shipped so far: an AI token gateway, the Trade Prospector foreign-trade platform, the VortoCode multi-agent development framework, plus two desktop tools — VortoShell and Vorto Write.
AI × Product × People: a full creation loop from idea validation to prototype to shipped iteration.

Skills

Machine & Deep Learning

My degree focus and the part I actually ship at work: dataset construction, fine-tuning, deployment and accuracy tuning.

PyTorch

Transformers

YOLOv5

BERT

LSTM

Writing

Moments

Snapshots of life & travel

Zhejiang Aisouls Intelligent Technology Co., Ltd.

2026.04 – Present

AI Engineer

Design and delivery of a multi-agent conversational system: LangGraph drives the main flow, with life-service business logic split into agent modules that can evolve independently.
Built out the real-time voice pipeline — streaming ASR/TTS with interruption handling — so web, terminal and hardware devices share one conversational interface.
Brought AI into community life services end to end: intent understanding, content generation, human review, and semantic-similarity matching between users.

Ningbo Qiee Longjoy Auto Parts Co., Ltd.

2025.11 – 2026.03

Full-stack Engineer

Built out the company's internal business workflow: steps that had lived in spreadsheets and verbal handoffs became an online process, with each stage's owner and transition rules made explicit and every step traceable.
Layered business agents on top of that workflow: repetitive information gathering and first-pass judgement run through an agent, leaving people to confirm at the decision points without giving up control of the process.
Delivered it full-stack — data modelling, backend APIs and frontend UI in one line — so a change to the process could land back in the system quickly.

Zhejiang Botai Linac Co., Ltd.

2025.01 – 2025.11

Full-stack Engineer

Replaced manual stocktaking on the production line with automated visual counting: built the dataset, fine-tuned YOLOv5 to run in real time within the compute budget of embedded hardware, then solved the hard part — double-counting across video frames — with a virtual counting line, reaching ~90% accuracy in production.
Handed product categorisation, until then a judgement call made by experienced staff, over to a model: fine-tuned bert-base-chinese on messy, non-standard product names to pass 96% accuracy on the test set, retiring the manual step entirely.
Joined client, order-batch, billing and payment tables into a single customer profile, segmented accounts into 9 value tiers with an RFM model, and layered churn early-warning on top — turning "who is worth chasing first" from a hunch into a quantified ranking.

happyding@home:~

            ~~~~~:~~·                       :~~~~~~~
          :+*$@@@@@$x:~                  ·:=@@@@@@$*::
        ··x$$$$$$$$$$$x+·              ·+o$$$$$$$$$$$x··
       ·:o%$$$@@@@@$$$$$:·            ·~@$$$$@@@@@$$$%o:·
       ·+$$$@@@@@@@@$$$$+~            ~:@$$$@@@@@@@$$$$+·
       ·+$$$@@@@@@@@@$$$$*··        ··o$$$$$@@@@@@@@$$$+·
       ·+$$$@@@@@@@@@@$$$*··        ··x$$$@@@@@@@@@@$$$+·
       ·+$$$$@@@@@@@@@$$$*··        ··x$$$@@@@@@@@@$$$$+·
        ~+%$$$@@@@@@@@$$$*··        ··x$$$@@@@@@@@$$$*+~
        ··*$$$$@@@@@@@@$$*··        ··x$$@@@@@@@@@$$$*·
        ··=x$$$@@@@@@@@$$*·=xxxxxxxx=·x$$@@@@@@@@$$$x=·
          ·~%@$$$@@@@@$$$$@$$$$$$$$$@%@$$$@@@@@$$$@%~·
           ··=@@$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$@@o··
           ·~+=%$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$%=+~·
          +o%$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$%o+
       ·:x@$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$@x:·
       ·+$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$+·
      ·o@$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$@o·
     ··x$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$*·
    ·~%$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$%~·
    ·~@$$$$$$$$$$$@$$$$$$$$$$$$$$$$$$$$$$$$@$$$$$$$$$$$$$@~·
    ·~@$$$$$$$$$@*oo+=%$$$$$$$$$$$$$$$$$$%xo++*@$$$$$$$$$@~·
    ·~@$$$$$$$$$x·%*··~@$$$$$$$$$$$$$$$$@·o%:··*$$$$$$$$$@~·
    ·~@$$$$$$$$$*·····~@$$$$$$%xx*$$$$$$@~·····*$$$$$$$$$@~·
     ·+*$$$$$$$$@%==+o@$$$$@@$%:~*$@@$$$$@o+==%$$$$$$$$$*+·
      ·o$$$$$$$$$$$$$$$$$$$@ooxxooxo%$$$$$$$$$$$$$$$$$$$o·
       ~=@$$$$$$$$$$$$$$$$$$$%@$$@%$$$$$$$$$$$$$$$$$$$@=~
        ·:x@$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$@x:·
          ·:o%@$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$@%o~·
             ~~ox%@@@$$$$$$$$@****%$$$$$$$$$@@%xo~~
               ·:+==+x*****%*:=o=+:o%******===++·
               ·:==========+~+%@xoo~:=====+====+·
              ·o%o==========~+oooo=~:=++=======*x·
             ·~*$$$$@*%***%%*=+=+=+x%******%$$$$%:·
            ·~%$$$$$$$$$$$$$$$$@@@$$$$$$$$$$$$$$$@:·
            ·:@$$$$$$$$ox@$$$$$$$$$$$$@*o@$$$$$$$$+·
           ·=$$$$$$$$$$@x+x@$$$$$$$$$*=o@$$$$$$$$$$o··
           ·o$$$$o*$$$$$$@·*$$$$$$$$@·x$$$$$$@o@$$$x··
           ·o$$$$*=o%$$$@o+@$$$$$$$$@==%$$$%x+x$$$$x··
           ·o$$$$$@*o=ooox@$$$$$$$$$$$*ooooo*@$$$$$x··
           ·o$$$$$$$$@@@@$$$$$$$$$$$$$$$@@@$$$$$$$$x··
            ~+@$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$=~
             ·=*$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$$%o~·
              ·=@@$@*xxxx*%$$$$$$$$$$%*xxxx*%$$@o·
               ~~xx=*@@@@%o=*%%%%%%%=o%@@@@%=ox~~
                 ~·+@$$$$$x··~~~~~~~·o$$$$$@=·~
                  ··:+++++··        ··:++++:··

Quantitative Trading System

Solo build

in usePython · FastAPI · React

Research → backtest → paper trade → review: turning an idea into a strategy that has actually been tested (A-shares / crypto / HK & US equities).

The backtest engine handles multiple timeframes and instruments with realistic market/limit fills, commission and slippage, reporting Sharpe, Sortino, Calmar and max drawdown.
One market-data interface across venues: Binance for crypto, akshare for A-shares and HK/US equities, switching between candles and live ticks; paper trading runs off the Binance WebSocket feed.
A React + ECharts frontend for equity curves, drawdown and return distributions, with cross-sectional factor screening and portfolio optimisation results compared side by side.

Python · FastAPI · React

Quantitative Trading System

Solo build

Keep scrolling →

DeskPet

Solo build

in betaSwift · SwiftUI · macOS

A pixel desk pet living in the macOS menu bar, watching your AI coding tools work.

Tracks 11 AI coding tools (Claude Code, Codex, Gemini CLI, Cursor and more): whether one is thinking, waiting on you or finished, the pet reacts instantly and notifies you — one click jumps back to the right terminal tab or editor window.
Approval requests from Claude Code and Codex land in a desktop action centre showing the command or file diff, answerable with allow-once or deny. Requests live in memory only, and anything unexpected falls back to each agent's native prompt — it never decides for you.
Project Islands lays out the last 90 days of work as a pixel panorama sized by activity: open one for cumulative sessions and per-tool share. GitHub repos also surface PRs awaiting your review and issues assigned to you — reusing your existing gh login, storing no credentials.

desktoppet.vortotech.com↗