About

Joe Tan

Fourteen years in military aviation. I built and run the development arm of the Air Force's drone unit, turning emerging AI and commercial compute into flying, mission-capable platforms — and nine years of operational air defence before that. I write the code as well as the programme.

Singapore Drone & AI programme lead Open to conversations Updated
ARGthe argument

Capability stopped being the hard part

A capable drone can be assembled from commercial parts by a small team in a few months. The industry has proved that at scale, and it is no longer where the difficulty lives.

What has not kept pace is the ability to say, with evidence, that the thing is safe — that the hundredth unit behaves like the first, and that it can be built at volume by people who did not design it. That gap is where programmes die. Not in the demo. In the step between a working prototype and a platform an operator will certify and depend on.

My work sits in that step: the standards, the test and evaluation cycle, and the requirements discipline that let a build survive contact with a real operator. Reproduction and scaling are the parts I claim expertise in.

It is also why the flight-test page on this site publishes the figures that look bad alongside the ones that look good. A number you cannot reproduce, and whose limits you cannot state, is not yet a result — it is a demo. See what that looks like in practice →

SVCthe day job

Fourteen years, two halves

In uniform since 2012. The work splits cleanly in two, and the first half is the reason the second one is any good.

Now · The Air Force

Development lead, the Air Force's drone unit

Built the unit's development arm from its stand-up — there was no R&D function before the role, and the technology-scouting-to-integration pipeline is mine. I direct a research team across computer vision, language models, edge inference, spectrum and robotics, then converge separate research lines into integrated flying platforms. I run test and evaluation end to end, to a standard the operational customer certifies against.

Alongside it, two applied-research builds inside the service: live speech-to-text transcription of air traffic control audio, feeding processed information into operational use; and an EEG-based brain-computer interface toward an operational use case.

Nine years before that · Ground-based air defence

Where the operational instinct comes from

Platoon command, battery operations, and chief training officer for a battalion — including training and certifying the service's short-range ground control intercept controllers. That is where the spectrum, comms and air-picture background comes from, and it is the reason I design to the operator rather than to the lab. Someone has to set the thing up in a field, at night, having not built it.

Defence work is described at capability level throughout. No platforms, no performance figures, no named partners, no operational detail. Further detail in conversation, where it belongs.

DEPTHwhat I am deep in

Where the hands-on actually is

Drones / UAS
Fixed-wing, quadcopter and helicopter — build, integrate, evaluate, operate, and build from scratch, including speccing the airframe to the mission. Autonomy stack: flight control, waypoint navigation, autonomous decision-making, swarm behaviour. Counter-UAS. CAAS UA Pilot Licence.
Test & evaluation
The full trial cycle — battery and motor bench tests, airframe integrity, endurance, through to customer certification. Writing the standard a build is measured against, not only running the trial.
Edge compute & AI
Jetson, Raspberry Pi and custom boards; deploying and integrating components to run models on-platform. Computer vision; small quantised models and LLMs assessed for feasibility at the edge; classical mathematical methods where no off-the-shelf model fit.
Sensors & spectrum
EO/IR, radar and RF comms across the consumer frequency range — how communication systems combine and behave across the spectrum, from an air surveillance background. Feeds straight into platform sizing.
Software
Python · TypeScript / Next.js / React · PostgreSQL / Supabase · C++ (embedded) · Docker · self-hosted LLM inference (llama.cpp, Ollama) · RAG · MATLAB
EDUeducation
Aug 2023 – Apr 2025

MSc, Artificial Intelligence

Nanyang Technological University, Singapore. Taken part-time, alongside full duty.

2013 – 2017

BS, Aeronautical Engineering

United States Air Force Academy, Colorado Springs.

BUILToutside work

Things I have built on my own time

Personal hardware, personal time, public data — which is exactly why they are here. These carry the numbers the day job cannot.

Vision-only approach guidance this site

A GPS-denied navigation and landing system for a Cessna 172, flown 216 times in one unattended 7-hour run with the decision-making on an embedded module. Measured envelope, published failure limits, every start reported. The whole campaign →

tcgTalk tcgtalk.com

A trading card marketplace for Southeast Asia, shipped and operated as sole engineer — auctions, fixed-price sales, trading, forum, collections: 238 API routes, 228 database migrations. Trained and shipped an anomaly-detection model that flags abnormal card price movements on live market data. Runs LLMs in production on self-hosted hardware with no third-party inference API, and operates unattended on nine scheduled jobs.

SmallTV firmware ESP8266 · C++

Constrained embedded firmware on real hardware: 413 KB in 39.6% of the app region, ~43 KB heap free. Serves its entire configuration web UI from flash with zero external assets, so the device stays manageable at a venue with no internet. Designed for honest failure — a stale reading greys out rather than showing a number that is no longer true.

drone_news_bot Python

Aggregates 20 drone and UAV sources, RSS-primary with scrape-fallback to survive bot blocking. Running unattended in production for a year. Also the reason I am never short of raw material.

Start a conversation

If you work on autonomy, GPS-denied navigation, edge inference, or test and evaluation, I would like to hear from you — whether that is a role, a collaboration, or a problem you are stuck on. I am open to conversations about what comes after fourteen years in uniform, and happy for the first one to be about the engineering rather than about me.

A resume is available on request.

Get in touch

This page describes personal work and service experience at capability level only: no employer systems, platforms, partners or performance data appear on it. The flight-test work published on this site is personal — personal hardware, personal time, public imagery and published airport data.