I build the systems companies run on — end to end, in production
0+Years Experience
0REST Endpoints
0KLines of Code
0Certifications
I am an IT systems and infrastructure engineer in Greater Vancouver. I started in 2022 as a
system administrator at a furniture manufacturer. Since 2025 I have been the sole engineer of
the ERP and automation platform two furniture manufacturers operate on daily — 274 tables
across four databases. I own the schema, the backend, the front end and the deploys.
My work is data modelling, idempotent integrations, and access control enforced at one
choke point. I build in Node, Express and MySQL, ship every screen in English and Chinese,
and apply language models where the job used to be re-typing. I am open to platform,
integration and internal tooling roles where one engineer owns the system.
What I Do
AI & Machine Learning
Function calling over live data, hybrid retrieval, IMAP commands, phishing detection.
System Administration
Windows/Linux servers, AWS, Google Workspace, network and endpoint hardening.
Full-Stack Development
Internal platforms and operator tooling on Node.js, Express and MySQL.
Cybersecurity
Security audits, penetration testing, DNS filtering, vulnerability scanning.
ERP & Automation
ERP and warehouse systems, accounting sync, EDI over SOAP, scheduled reporting.
One platform, two companies — designed, built and run by one engineer
What it is: one website and one login where staff take an order,
check whether the piece is on the warehouse floor or still on a ship, print the
shipping and customs paperwork, and hand the figures to accounting without anyone
re-typing them. Behind it: one Node/Express codebase, one permission model,
77 route files, 24 services, bilingual EN/ZH. The primary database sits
overseas, a few hundred milliseconds round-trip, so the real work is query batching,
a local read mirror and caching rather than CRUD.
1,048REST Endpoints
274DB Tables
230KLines of Code
124Permission Checks
Node.jsExpressMySQL (4 DBs)RBACOpenAI Function CallingIMAP / NodemailerSOAP / EDIPuppeteeri18n (EN/ZH)PM2
The sign-in screen — the only one I can publish. Everything past it reads live company data.
Try It
Three runnable panels — an assistant sandbox, a recorded search rebuild, a caption planner.
AmaxBot · interactive
How it picks a tool, calls it and answers
Each run shows the routing decision, the tool call with its real
argument shape, and the rows returned.
This is a sandbox. It runs in your browser on static sample
data — fictional customers, no network calls, no database. The real assistant sits
behind a login and reads live company data.
amaxbot — sandbox sessionsample data
Nothing has run yet
Pick a question above, or type your own below. Start here:
Free text is keyword-matched to the nearest of the eight sample questions, and
the panel says so plainly when it cannot find one.
Recorded run · 2 Sep 2026
Keyword search vs. hybrid search
The knowledge base behind AmaxBot was searched with MariaDB FULLTEXT.
Chinese was unsearchable and paraphrased English missed. I shipped a hybrid
embeddings + FULLTEXT layer and ran both paths against the same
eight staff questions.
This is a replay of one measured run, not a live query.
Every question and article title below is a verbatim record of what the two
search paths returned. The knowledge base holds customer data, so nothing
on this page talks to it.
5/8Keyword search returned any rows at all
8/8Hybrid search returned the correct article
0/3Chinese questions keyword search could answer
Not the same measure. Returned any rows is a low bar: on one of the five the rows
were the wrong article. The hybrid column counts only the correct article.
Pick a question from the run
Five of the eight questions had their result pairs recorded verbatim; those five are
the ones shown. The other three were English questions both paths returned rows for.
Keyword search
MariaDB FULLTEXT
Before
Hybrid search
Embeddings + FULLTEXT
After
26 published articles110 chunks0.64 MB of vectors$0.0005 per full rebuild
Social Video Maker · interactive
Where the pauses go
The free speech engine returns zero bytes for any SSML pause tag, so the pacing is
built in reverse: split the script, synthesize each fragment on its own,
and splice measured silence between them. Type below and watch the plan change.
This is the real pacing rule, running here. The gap lengths are the
constants from the renderer. The browser cannot synthesize the audio, so the duration
here is an estimate; the renderer measures the real thing.
380 ms after a sentence180 ms after a clause500 ms at a paragraph120 ms at a line break
Case Studies
Four builds, each written as problem, mechanism, result.
AmaxBot — a GPT assistant that inherits the ERP’s permission model
AI / Node
Problem
Where is this order, is it in stock, what does this account owe — every answer sat behind SQL across four databases and 274 tables. Staff asked IT; IT answered hours later with a screenshot.
Built
Two surfaces over one accuracy layer: a chat UI exposing 55 function-calling tools, and an IMAP service answering 27 @amaxbot_* commands from a shared mailbox, each in English and Chinese.
Result
Staff query the ERP themselves, by chat or plain email. The tool catalog is filtered by role before the model sees it, so a non-admin request cannot contain a write tool; out-of-scope rows never reach the model.
The integration layer — EDI in, accounting out
SOAP / Accounting
Problem
Purchase orders arrived from trading partners over EDI, then the same orders and supplier bills were re-keyed by hand into two accounting companies. Nobody caught a transposed figure until month-end close.
Built
Inbound, a SOAP client against an EDI provider: pull partner purchase orders, return acknowledgment, ship notice and invoice. Outbound, an accounting hub of 6 connectors and 42 APIs posting sales documents and supplier bills on a daily schedule.
Result
Partner orders reach the ERP and both accounting companies without re-keying. Every push is idempotent, keyed on the document number, so a resend cannot double-post; a rejected batch is un-flagged and retried.
Knowledge-base search — rebuilt as hybrid retrieval
Embeddings / RRF
Problem
The knowledge base ran on one MariaDB FULLTEXT match — literal words only. Chinese failed outright: no CJK tokenizer, so a whole question arrives as one unmatchable token, and the Chinese columns sat outside the index.
Built
A retrieval layer embedding 26 articles as 110 chunks, fused with the existing FULLTEXT query by reciprocal-rank fusion. No vector database: 0.64 MB of Float32 blobs, cosine-scored as a dot product in Node.
Result
Five of eight staff questions returned rows before, one of them the wrong article; all eight now return the correct article, including three Chinese questions that returned nothing. It ships behind a flag and falls back to the old query.
Social Video Maker — script to vertical video
FFmpeg / TTS
Problem
Marketing needed a steady stream of Mandarin short-form video with no per-video budget and no editor in the loop: a publishable 1080×1920 file produced from pasted script text in a single pass.
Built
A standalone Node app: TTS narration, Ken Burns b-roll, a talking-head overlay driven by per-100 ms amplitude samples, burned-in captions, FFmpeg assembly, and progress streamed live to the browser. Every service in the chain is free.
Result
The free speech endpoint returns zero bytes for SSML breaks, so pacing is spliced silence and each caption boundary snaps to the nearest measured pause. Publish-ready in one pass, covered by a 130-case unit suite.
Other Projects
Email Phishing Detection
A Google Apps Script on a timed trigger submits links and attachment hashes to a reputation service, then a language model reads the body for pretext.
Apps ScriptMLChatGPT APIVirusTotal
Voice Assistant
A Python desktop tool that transcribes speech in real time with Vosk running on-device, then answers questions against the running transcript.
VoskChatGPT APIPython GUIAutomation
EDI & API Integration
Order and stock feeds between the ERP, marketplace back ends and warehouse systems, normalized to one order shape, upserts keyed on the partner’s document number.
EDIREST APIAutomation
Network Monitoring
Scheduled probes collect link, latency and service-availability metrics into a time series, rendered as Grafana dashboards a non-engineer can read at a glance.
NetworkingGrafanaPython
Python Automation
Walks a folder of supplier PDFs and spreadsheets, extracts the tables, reconciles them against the ERP and writes an Excel exception workbook.
PythonPDFExcelAutomation
Desktop Chat Assistant
A Python desktop app that watches an instant-messaging window, reads the thread with a vision model and holds each drafted reply for human approval.
PythonClaude APIVisionAutomation
CCTV Archiver
Pulls recordings off a Hikvision NVR on a schedule and uploads them to S3 through presigned PUT URLs, so the device holds no AWS credentials.
PythonAWSAutomation
Bilingual News Digest
A Node service on a timezone-pinned cron that gathers industry news, summarizes each story in English and Chinese, and mails a rendered HTML digest.
Node.jsCronChatGPT APIi18n
Browser Automation
Playwright scripts that drive a saved browser session to reach partner portals and upload the inventory feed, formerly posted by hand, roughly 177 rows a run.
Major GPA 3.8 — web development, IoT, cloud security. CO-OP completed.
Web DevIoTCloud Security
Work Experience
Current
May 2025 – Present
IT Systems & Infrastructure Specialist
Amax Leather Canada Ltd & Aurora Furniture MFG
Sole engineer of the ERP and automation platform both companies run on: 1,048 REST endpoints, 274 MySQL tables, 150 pages, ~230,000 lines in production — plus the in-house GPT assistant, EDI and accounting integrations, and the dealer portal.