AI Application Engineer · Agent Platforms · Systems Integration

Ken Huang黃鵬展

Moving AI from “can answer” to “reliably gets the job done.”

14 years in software, including 11 years owning Murex FX-trading and STP systems at Taishin Bank. Lately I focus on AI agents, MCP tooling, and model training and evaluation; PickPico, built with teammates, took third place at FUTUREMODE × SITCON Hackathon 2026.

in software, 11 on trading systems
14yrs
counterparty-bank STP onboardings
7
FUTUREMODE × SITCON Hackathon 2026
3rd
projects built, public and private
16

How I work with AI

How I work with AI

I own the scope and acceptance bar, let agents carry the implementation, and let real devices and data decide.

  1. 01

    Define scope and acceptance first

    Before writing code, I pin down who operates it, what must work, and which settings and data must survive. Cheap, reversible changes come first so the results steer the next move.

    ExampleFor my RL agent, acceptance meant win rate over blocks of live fights—never one lucky fight.

  2. 02

    Let agents carry the implementation

    I work with Claude Code, Codex, and Gemini on implementation, analysis, or review depending on the task, turning recurring workflows into tools or skills.

    ExampleOn PicoBridge, agents wrote much of the cross-stack code across a Windows driver, C# desktop app, and Android, while I focused on Bluetooth standby and clipboard behavior.

  3. 03

    Verify on real devices and data

    A command returning OK, a clean install, and a tool that feels right in daily use are three different milestones. I test on my own phone, head unit, peripherals, and live games—if it stalls in practice, it isn’t done.

    ExamplePicoCar reported success setting the default launcher, yet the physical HOME button still opened the factory UI. I tracked down the vendor setting, fixed it, and tested the actual button.

Featured

Agent platforms and model evaluation

All 16 projects

Experience

Experience

14 years in software, 11 of them on financial trading systems.

2015/05 – 2026/08

Taishin International Bank

Senior Software Engineer | FX Trading · Murex · STP

11 yrs 4 mos

Owned limits, workflows, and pre-trade rules in the Murex FX platform, and designed and ran the STP server between Murex and counterparty banks.

  • Configured and maintained Murex MLC (market limit control), Exchange Workflow, and Pre-trade Rules across front, middle, and back office.
  • Built data-transformation and field-mapping logic in Murex Formula (XSL/XSLT); maintained the Confirmation module.
  • Designed and ran a standalone C#/VB.NET STP server, monitoring interface logs, trade flows, and performance to cut latency, data mismatches, and dropped deals.
  • Onboarded STP for seven counterparties: EBS, Reuters, UBS, S&P Global, Deutsche Bank, JP, and SPARK.
  • 2024–2026 bank merger: wrote the Murex-to-NX data-conversion and interface formulas.
  • Packaged repetitive test documents and procedures into AI-agent skills for colleagues, cutting repetitive form work by about 80%.
MurexMLCSTPXSL / XSLTC#VB.NETSQLOracleSQL ServerMQSocketSFTPWindows ServerLinux
2012/09 – 2015/04

Kewei Information

Software Engineer / FAE Team Lead

2 yrs 8 mos

Backend development for a travel-information platform, plus leading the FAE team on production issues.

  • Built and maintained backend features for a travel-information website; debugged application logic, databases, and systems.
  • Led a three-person FAE team: task allocation, technical support, and coordination.
Classic ASPSQLProduction Support

More projects

More projects

Public and private, and most of it is something I use every day.

ML & Evaluation2026

RL Boss-Fight Agent

An agent fights the Orphan of Kos in Bloodborne at native speed, pairing a PPO policy with fast reflexes and human rules—35 wins in 44 live fights on the retained build. Later experiments separate observations, rules, and learning to push more decisions into the policy.

Reinforcement LearningLive-tested
PythonPyTorchStable-Baselines3MaskablePPO
Details

Agents & MCP2026

PicoCar

A custom launcher and AI assistant for my Android head unit, unifying navigation, media, shortcuts, and app windows with shared operations for the local assistant and external MCP. Installed on an Android 13 car head unit, with physical HOME verified to return to PicoCar.

Real Head-unit InstallIn Development
AndroidJavaMCPAccessibility Service
Details

Agents & MCP2026

MCP Relay & Camera Tools

One Cloudflare Worker plus a Windows bridge that safely exposes several local MCP servers to remote AI; its camera MCP lets an agent see, pan and tilt, listen, and speak.

MCP InfrastructureDaily Use
MCPCloudflare WorkersDurable ObjectsOAuth
Details

Agents & MCP2026

AnkiTutor MCP

An MCP server turning any chat client into a language tutor: Anki handles scheduling and decks while the AI teaches and grades. I use it daily through a ChatGPT connector.

MCPDaily Use
PythonMCPAnkiConnectpytest
Details

Agents & MCP2025–2026

coding-tools MCP

My local AI-collaboration layer: cloud chat AIs work inside approved local workspaces, shells, Git, and connected devices, and pass the smallest needed step to me via human_help_me when stuck.

Internal ToolDaily Use
MCPPythonShellGit
Details

Agents & MCP2026

Paper2AI

Started from a stack of health-report photos as a page-by-page transcription tool into Markdown with page numbers and source links, then grew into a scanning workspace with crop, flattening, AI OCR, side-by-side proofing, and PDF/Markdown export.

PWAPrivate
JavaScriptPWANode.jsVision LLM
Details

ML & Evaluation2026

DrumLab & music-sep

My own drum-practice workstation: Demucs splits stems, ADTOF transcribes drums to MIDI, and a browser editor lets me record my e-drum kit bar by bar and export standard MIDI.

Audio MLDaily Use
PythonPyTorchDemucsADTOF
Details

Systems & Devices2026

VibeDeck

Turns spare phones, tablets, and e-readers into real Windows extended displays or crisp info decks—complete with system telemetry, AI-quota boards, and remote control from a phone.

Public Release
C# / .NET 8WebRTCWindows IDDPWA
Details

Systems & Devices2026

PicoBridge

Lets my desk Windows keyboard and mouse switch smoothly to phones and a second PC while carrying clipboard and files along. Grew from an ESP32-S3 hardware bridge into a native Windows Bluetooth HID driver, desktop app, and Android companion built for daily switching.

Daily UsePrivate
C#C (Windows driver)JavaAndroid
Details

Systems & Devices2026

AeroDoc

Turns any local folder into a lightweight document workspace: one desktop app reads Markdown, PDF, DOCX, XLSX, and CSV with search, notes, and light editing—files stay right where they are.

Open SourceMIT
Tauri 2RustReactTypeScript
Details

Finance & Research2026

Equity Lab

Tracks large-holder shifts in Taiwan stocks using TDCC shareholder-distribution data, generating rule-based signals and event backtests—intentionally skipping ML in the core so every signal stays traceable and reproducible.

ResearchPrivate
PythonFastAPIPolarsDuckDB
Details

Finance & Research2026

BacktestLab

Compares lump-sum, monthly DCA, and monthly-plus-dip-buying on equal starting capital—every purchase is auditable, uninvested cash counts toward returns, and price vs. adjusted close is flagged clearly.

Open Source
Node.jsJavaScriptnode:test
Details

Finance & Research2026

Cast Iron

My private investment ledger: ingests brokerage records daily, keeps complete trade history, and calculates performance and multi-year dividends under my own accounting rules.

Daily UsePrivate
RustCrawlerEmail ingestionReporting
Details

About

About

Trading systems cannot afford dropped deals or data mismatches. A decade of Murex, STP, and counterparty-bank integration trained me to pin down data flows and failure paths upfront, and hunt down root causes in production.

I hold AI to the same bar: models have to plug into real tools and hardware, and their output has to hold up in testing. AI coding agents carry most of the implementation; I own the scope, trade-offs, and real-device sign-off.

AI & Agents

AI AgentLLMMCP / Tool CallingHuman-in-the-loopLLM-as-judgePlaywright

Machine Learning

PyTorchReinforcement Learning (PPO)Behavior Cloning / DAggerLoRA fine-tuningCalibrationStable-Baselines3Demucs

Languages

PythonC#VB.NETTypeScriptJavaScriptRustJava (Android)PowerShellSQLXSL / XSLT

Fintech

MurexMLCSTPPost-Trade ProcessingCounterparty integration

Platforms & Tools

AndroidCloudflare Workers / Durable ObjectsTauriWebRTCDockerSQLiteDuckDBWindows IDDPWAGit

AI Dev Tools

Claude CodeOpenAI CodexGoogle GeminiChatGPT

Education

Education

  • National Taipei College of BusinessInformation Management, two-year technical program · 2009 – 2011
  • Taipei College of BusinessInformation Management, five-year junior college · 2003 – 2008

Certifications

Certifications & awards

  • Microsoft Certified Professional (MCP)
  • FUTUREMODE × SITCON Hackathon 2026, Third Place (PickPico)

Languages

Languages

  • Mandarin: native
  • Taiwanese: conversational
  • English: technical reading
  • Japanese: basic

Contact

If this fits your team, write to me directly.