Tony Huang, portfolio

Sydney, New South Wales

TonyHuang

Most engineers show you what they built. I will also show you how I stopped it going wrong.

35,000+
Automated writes to the live property portal and Dropbox, with every portal listing change recorded to an audit log.
102
Of 161 human review decisions were rejections, which is what makes the approval gate real.
803
Commits on that platform, all of them mine, over three and a half months.

AI and Automation Engineer. Data and Analytics.

Master of Information Technology (Data Analytics), UTS

AI and Automation Lead, VPI Group. Four days per week on site.

Work rights

Full Australian working rights. Temporary Graduate visa (subclass 485), valid to 2028. No employer sponsorship required.

01

What I do

I take an ambiguous business complaint and turn it into a system people use every day.

I am the internal AI and automation lead at a Sydney property investment group. For three years I have been translating high friction manual business processes into automated systems, and since early 2026 I have been designing and delivering production generative AI applications end to end: scoping with non technical stakeholders, solution design, build, production hardening and user adoption.

I work alongside sales, accounting and property management. I find the workflows costing the business time or accuracy, gather requirements from people who do not speak in tickets, design the solution, build it, and own it through to production and adoption. The lead role was not assigned to me first: I built a small tool off my own initiative, the owner saw what it did, and the company's first internal AI and automation lead role followed from that.

02

Selected work

Three systems built end to end at a single employer, VPI Group (Vision Property Investment), and one public repository you can run yourself.

01

VPI Hub, the platform that publishes to the live portal

In production, unattended

Python, FastAPI, server rendered templates with Alpine.js and Tailwind, SQLite, APScheduler, Playwright, a vision model used only to read unit numbers off plans.

Doing the same coverage by hand takes about 123 hours, which is why nobody had ever done it completely. This did not make the job faster. It made an undoable job happen every week.

Over twenty developer sources arrive as price lists and floor plans, in whatever shape each developer likes. The platform collects them unattended, extracts them, matches them against about 250 projects on the company's live public property portal, and writes back through a human approval gate with an audit log behind it. Twenty page routes and 152 API routes, and the people who use it are not engineers: they drag schedules on a calendar, approve or reject in a queue, and edit project details that write back to the portal.

  • The hard part is dirty input, not architecture. Every developer's spreadsheet is laid out differently, and the same unit is written one way on the price list and another way on the floor plan. Matching runs on normalisation and fuzzy linkage with confidence bands, and anything under the automatic threshold goes to a person.
  • A write window that is only ever open for one project. Each production write is fenced to the single project being processed and closes immediately after. If a file cannot be read it is held rather than guessed at, a file with an anomalous page count is refused, and if the vision quota runs out the system will not fall back to the weaker reader. Output identical to the input while claiming to have marked up N units is a hard error.
  • Multi source reconciliation found problems nobody had reported. Comparing the price list, the marketing floorplates, the archived files and the live portal surfaced seven units carrying another building's price, and sold stock still publicly listed. The drift came from handoffs between systems rather than careless typing, so the fix was a scheduled reconciliation rather than a one off cleanup.
  • I found a concurrency fault by reading a broken artefact. A scheduled job and a hand run job reached the markup stage through different entry points, shared no lock, and produced a zero byte deliverable. All paths now take one operating system level file lock.
35,000+
automated writes to the live portal and Dropbox: listings created, repriced, withdrawn and relisted, plus price lists and floor plans filed
24
developer sources collected unattended, reconciled against about 250 projects on the live portal
2,359
price labels printed onto 505 floor plates, held back rather than published when a check fails
2,687
test functions across 210 test files, a test to code line ratio of 0.71
803
commits, all mine, 12 May to 26 August 2026

What it is not, because the scale invites the wrong guess: a single tenant internal platform on one machine with one SQLite database and an in process scheduler. Not multi tenant, not distributed, not microservices. It holds developer credentials and full pricing, so it is deliberately not opened up to the sales floor. Running it on a server with CI is the next thing it needs.

02

Agentic AI voice assistant, in the company CRM

Feature complete, pending compliance sign off

ASP.NET Core 8, .NET MAUI (Android), Google Gemini 2.5 Flash, Azure SQL, JWT authentication.

Sales agents speak or type a question in English, Chinese or a mix of both, and the assistant either answers it or files the follow up note against the client record. It covers property search, affordability and loan calculations, and client log updates in one chat surface on the mobile app, replacing thumb typing in the field and switching between five separate screens. It is feature complete, pending compliance sign off and production authentication cutover.

  • A tool calling agent, not similarity search. For a sales tool quoting prices and availability, similarity is the wrong primitive. The data has to be exact and live, so the model picks the tool and fills the parameters, and SQL does the rest.
  • Intent recognition and slot filling. The language model turns natural language into typed search criteria, then selects from roughly 12 deterministic tools that query Azure SQL. The model never generates a figure itself. Speech to text runs multimodal through Gemini with schema constrained output.
  • Bilingual by design. English, Chinese, or a mix of both inside the same sentence.

03

Auto Tenant Check, and the market report generator

In production, adopted daily

Playwright browser automation, Vision OCR, GraphQL API integration, Python, LLM report generation, installable Windows application.

Auto Tenant Check is a multi stage tenant screening tool integrating email intake, browser automation with Playwright, Vision OCR and credit bureau data. I wrote the pass mark into the tool before measuring anything: at least 30 comparisons, and 95 percent on the fields that gate a decision. Benchmarked against 30 real applications it came out at 95.3 percent across the five gating fields, with phone and both address fields perfect but email at 80 percent, so email failed the bar I had set. Rather than tune the reader I diagnosed the ceiling as watermarking, reverse engineered the upstream GraphQL API, and moved nine text fields onto the structured source in a staged shadow then cutover migration. That does not improve the transcription error, it removes it by construction. Identity document numbers still go through vision extraction and I have no number worth quoting for them. It ships as an installable Windows application.

The market report generator is an LLM backed tool that lets non technical staff produce a polished property market report themselves. I gathered the requirements from the sales team and iterated on output quality from their feedback, and a non technical colleague now reaches for it by choice. That voluntary use is the part I am proudest of, because internal tools usually die on the vine.

github.com/Heatescape/cma_editor

04

plumbline, a reconciliation stage you can run yourself

Public, runnable, clean room

Python, PyMuPDF, pandas, pytest, GitHub Actions CI.

The work above is closed source and holds client data, which leaves a reviewer taking my word for it. So I rebuilt one stage of it from scratch against generated data: take a marketing floorplate PDF, reconcile it against the authoritative price list, and mark up what has drifted. Prices that moved get struck and rewritten, stock the list no longer carries gets a SOLD stamp, a label printed on the plate that does not exist in the list at all gets boxed.

  • It runs with nothing on hand. The tool generates its own synthetic inputs from a seeded random number generator, so reconcile demo goes end to end on a clean checkout. No real development, address, price, person or company appears anywhere in it.
  • The demo is part of the build. CI does not only run the 36 tests, it runs the demo and asserts the artefacts came out non empty. For a repository whose whole claim is that you can run it, a broken demo is a broken build.
  • It is a keyhole, not the platform. It deliberately covers one stage rather than imitating the whole system, because something built to look like a platform in a portfolio ends up looking like a toy in every part.
github.com/Heatescape/plumbline
03

How I ship safely

Publishing to a public property portal means a bad write is visible to customers.

So the controls came first, and they are the part of the work I would want reviewed hardest.

Traceability

Audit logging on every listing change

Every listing change on the live portal is recorded to an audit log and remains traceable after the fact.

Human in the loop

Risk based approval routing

Actions are routed by risk. Low risk publishes automatically. High risk goes to a human review queue.

Concurrency safety

One lock across every path that writes

Two entry points once reached the markup stage without a shared lock and produced a zero byte deliverable. Diagnosing that is what put an operating system level file lock on every path, rather than a lock on the path I happened to be looking at.

Confidentiality

Sensitive commission data filtering

Confidential commission figures are kept off the public portal, alongside versioned compliance archiving of every published price list.

Privilege separation

The model is kept out of the permission layer

In the CRM assistant, the identity used for a query is set by the authenticated session and can never be supplied by the model, so it cannot query on behalf of another user.

Bounded agency

A capped agent loop with full tool call tracing

The agent loop runs to a default ceiling of 6 tool calls, and every call is traced with its arguments and a result summary for after the fact review. The ceiling is a default rather than something the server clamps, which is worth saying plainly.

Anti hallucination

A faithfulness constraint in the data contract

The assistant must state when no option matches the buyer's budget before listing near misses, rather than relying on prompt instructions to keep it honest.

Regression testing

2,687 test functions on the portal platform

Across 210 test files, a test to code line ratio of 0.71, including regression suites built from real sample PDFs so that supporting a new developer format cannot quietly break an old one.

04

Skills

Named for what they actually are, not padded.

Artificial intelligence and LLM

Generative AI, large language models, agentic AI, tool calling and function calling, prompt engineering, intent recognition and slot filling, LLM document extraction, multimodal speech to text, structured output with JSON schema, anti hallucination guardrails, OpenAI GPT-4o and Vision, Anthropic Claude, Google Gemini, GitHub Copilot.

Data and analytics

Python (pandas, NumPy), SQL, Azure SQL, ETL pipeline design, data reconciliation, data quality and completeness scoring, fuzzy matching and record linkage, data validation, Excel.

Automation and delivery

Business process analysis, requirements gathering, stakeholder engagement, end to end delivery, workflow automation, browser automation with Playwright, Optical Character Recognition, scheduled unattended operation, user training and documentation.

Governance and reliability

Human in the loop approval workflows, audit logging and traceability, risk based approval routing, role based access control, idempotent operations, concurrency safety, regression testing, sensitive data filtering.

Platforms and integration

REST and GraphQL API integration, OAuth2, JWT authentication, FastAPI, ASP.NET Core, .NET MAUI, Azure Blob Storage, Dropbox API, SQLite, Git.

05

Background

The generative AI work sits on two and a half years of unglamorous process automation.

VPI Group (Vision Property Investment), AI and Automation Lead

May 2023 to present

Four days per week on site, alongside a Master of Information Technology.

Excel, VBA and Python automation

2023 to 2025

Before the generative AI work, I spent two and a half years automating the recurring reporting and data maintenance load across sales, accounting and property management, including the weekly developer price list update that later grew into the platform above. Unglamorous work, and it is where the process automation instinct and the data validation discipline came from.

Master of Information Technology (Data Analytics), University of Technology Sydney

Graduated July 2026

Bachelor of Information Technology (Data Analytics), University of Technology Sydney, 2024.

Building Permit Data Analysis, academic project

Python, individual

Cleaned and analysed a 198,900 record San Francisco building permits dataset with 43 attributes. Imputed missing values using median and mode statistics alongside a linear regression cost model, engineered derived date features to measure permit approval cycle times, and visualised permit type composition, seasonal application trends and variance between estimated and final construction costs. Tools: pandas, NumPy, scikit-learn, Matplotlib, Seaborn.

Certifications

Google

Google Data Analytics Certificate. Google Finance Data Analyst Certificate.

06

Get in touch

If you have a process that is quietly costing your team a day a week, that is the conversation I like having.