Alison TeohTeoh Yi Huan

Spinector

A digital scoliometer that measures spinal curvature using a phone's sensors. Version 1 made measurements more accurate and trackable. Version 2 added features for patients in rural areas, and version 3 added two-way communication between doctors and patients.

My role: backend developer. One line on what you built

Gold · GirlsCode Mobile Apps Hackathon 2022 (state) · Silver · Sarawak Mobile Apps Hackathon 2022 (state) · Silver · Young Innovators Challenge 2020 (national)

• MIT App Inventor
• Phone sensors
• Healthcare

How Much Wood Can a Woodchuck Chuck?

An interactive heatmap with a year slider that forecasts woodchuck "wood-chucking" hotspots across Pennsylvania for 500 years, plus a chuck graph and literal log rain. It's built from GBIF sightings, Pennsylvania Game Commission harvest reports, census population density and USDA wood-density data.

With only 2–7 observations per location, time-series models like ARIMA were out. We tried a random forest, then replaced it with an exponential growth projection with stochastic variation, and treated the uncertainty as part of the story.

My role: e.g. data cleaning in Pandas and the Polars projection

• Python
• Pandas
• Polars
• JavaScript

Flight Computer PCB

I designed the society's central flight-computer board, bringing GPS, an IMU and airspeed sensing together with a long-range RF control interface.

I also introduced the team's first formal wiring architecture to reduce EMI susceptibility and the risk of mechanical failure, debugged and wired electronics on test aircraft, and built an RC foam glider.

Optional: the MCU and radio you chose, and whether it has flown

• PCB design
• GPS · IMU · airspeed
• RF control
• Wiring architecture

Sensor-Communication PCB

Within a month of joining, I was chosen as lead designer for a custom board that carries sensor data around the pod. I owned it from schematic capture through to manufacture.

The board's I²C, CAN and SPI interfaces were designed to meet a sub-1 ms sensor-feedback requirement.

• Altium
• I²C
• CAN
• SPI
• Schematic → manufacture

DigDoc

A retrieval-augmented generation platform. It ingests PDFs and splits them into chunks, retrieves the passages relevant to a question, and grounds the LLM's answer in those source passages.

• Python
• LangChain
• RAG
• LLMs

Marketing Research MCP Server

An MCP server that hands Claude a research toolkit: scrape_page returns readable text with navigation and script noise stripped out, extract_meta pulls the title, description, canonical URL and Open Graph tags, analyze_headers maps the H1–H3 hierarchy, and compare_competitors fetches several sites at once with asyncio.gather and returns a comparison table, reporting any site that fails in its own row.

It also ships a prompt that chains those tools into a positioning brief, and a resource holding that brief's template. Written against the MCP Python SDK 2.x, whose API differs from the 1.x examples most tutorials still use.

• Python
• MCP
• httpx
• BeautifulSoup
• asyncio

Local RAG API

A FastAPI service that stores documents and answers questions from them. A question is embedded with nomic-embed-text, ChromaDB returns the closest chunks, and a local LLM answers using only that context. Every answer comes back with the passages it used, so you can check it.

Documents are chunked by paragraph and tagged with their owner, so a query can be filtered to one person's profile or search them all. Everything runs locally through Ollama: no API keys, and no data leaving the machine.

• Python
• FastAPI
• ChromaDB
• Ollama
• RAG

Oops, There's a Problem!

A Manifest V3 Chrome extension for developers about to launch a site. It checks every visible element against five rules: WCAG AA colour contrast (using relative luminance), missing or too-short alt text, unlabelled form controls, videos without captions, and interactive elements you can't reach with Tab. Each problem gets a red outline and a tooltip, and the toolbar badge counts them.

My role: e.g. wrote the colour-contrast checker

• JavaScript
• Chrome Extensions
• WCAG 2.1

Bayesian Coin Flip Simulator

Bayes' theorem built from the ground up in Python. A hidden coin, either fair or 60% heads, is flipped repeatedly, and the belief in each coin is updated after every flip. Across 200 simulations of 100 flips, the model picked the right coin 83.5% of the time, but reached 95% confidence in fewer than half the runs. Coins that behave similarly need far more evidence to tell apart.

• Python
• Probability
• Monte Carlo
• Matplotlib