Adrian Chen

Machine learning and AI engineering.

BSc Mathematics and Statistics at UCL, graduating 2027. I build machine learning systems, from research prototypes to production deployment.

Scroll

Selected work

Three projects, and the decisions behind them.

FYLD · 2026

On-device roadworks detection

A five-class roadworks detector running on Android. It gives immediate feedback on site and sends only the frames that need further analysis to a server-side vision-language model. I was responsible for the R&D from dataset to on-device validation. The signed-off model reached 0.5578 mAP50-95 at a sustained 4.04 fps.

On-device CV
Pseudo-label distillation
Thermal budget

Read more Read less

Problem

Running a large vision-language model on every frame is costly, slow and subject to rate limits. Field workers needed results within seconds.

Approach

The work is split by cost. A small on-device detector identifies which equipment is present and where, such as cones, barriers and temporary signage, and gives immediate feedback. Only frames that need further reasoning are sent to the server-side model, which reduces backend requests. I was responsible for the data, model selection, pseudo-label distillation, validation and on-device implementation.

Model selection

I compared the transformer-based RF-DETR with YOLO. RF-DETR Nano was the most accurate model trained, at 0.5997 mAP50-95, but at 30.16M parameters it was roughly twelve times the size of the deployable model. YOLO26n fitted the device budget at lower accuracy.

RF-DETR was therefore used as a teacher. It generated pseudo-labels on unlabelled footage, which were used to train the YOLO26n student. RF-DETR remains an option for higher-end devices. Licensing was also a factor in the choice.

Results and validation

The model was signed off on 2 September 2026.

Pseudo-label distillation improved YOLO26n by +0.0366 mAP50-95, with no added inference cost.
The gain held in all 5 cross-validation folds, at 2.15× the paired noise floor.
A budget-matched control, with the same compute and no additional data, changed by −0.002. This rules out longer training as the cause.
The signed-off model reached 0.5578 mAP50-95, with presence F1 of 0.82–0.94 across the five classes.

Thermal constraints

On a phone, sustained operation is limited by heat as well as accuracy. Thermal profiling showed that a PyTorch benchmark had overstated RF-DETR latency by 6.2× relative to ONNX Runtime, which changed the basis of the architecture decision.

Running on two CPU threads sustained 4.04 fps over a 30-minute soak test and lowered peak skin temperature by 3.2 °C compared with four threads. The handset still reached 57.1 °C and throttled, so overheating was reduced but not resolved.

Data

I built a model-in-the-loop labelling pipeline. A vision-language model generated initial labels, SAM refined them, and iterative self-training improved them further. After human review, the approved dataset contained 1,400 images and 3,530 boxes from 633 source videos, including 264 true negatives. I also identified a dependency licensing issue before deployment.

Stack

YOLO26n · RF-DETR · SAM · pseudo-label distillation · k-fold cross-validation · on-device Android inference

Signed-off mAP50-95: 0.5578
Gain from distillation: +0.0366
Sustained on device: 4.04 fps
Peak skin temp: 57.1 °C

FYLD · 2026

Job duration prediction

A gradient-boosted model for predicting field job duration. Organisation ID was the strongest predictor but gave managers nothing to act on. I reframed the objective around factors they can control, and kept predictions usable for customers with no history.

LightGBM
SHAP
Cold start

Read more Read less

Reframing the objective

The model used several years of operational data, including live updates, photographs, hazards and weather. Gradient-boosted trees suited the largely categorical features.

Organisation ID was the strongest predictor: the company doing the work was more informative than the work itself. Managers cannot change it, though. I shifted the objective towards factors they can act on, including crew composition, weather and access conditions. I used SHAP attribution to explain individual predictions.

Cold start

I limited how much the model could rely on customer history, so that new customers get usable predictions from the start. This reduced offline accuracy. Comparable jobs at similar locations provided an additional signal that does not depend on customer history.

University of Technology Sydney · 2025

Regulatory document retrieval

A retrieval assistant for pharmaceutical regulatory review, with citations at paragraph level. I tuned retrieval for recall, because a missed document costs more than an irrelevant one. Recall@10 was 0.9640 in a twelve-document proof of concept.

RAG
Two-stage retrieval
Evaluation

Read more Read less

Context

The Australian PBAC archive contains more than 2,000 Public Summary Documents. Evaluators have five to nine weeks to assess a submission, and much of that time goes on finding relevant precedent. The proof of concept was deployed on a twelve-document subset for technical validation, with answers linked to source paragraphs.

Recall over precision

Returning an irrelevant document costs a reviewer time, but missing a relevant one can undermine the review. I tuned the two-stage retrieval for recall and accepted lower precision to reduce the risk of omission.

Evaluation

Against a full-context baseline, the system scored higher on both lexical and semantic overlap. Against a prompt-guided baseline, it scored lower on ROUGE-1 but higher on BERTScore. Citation constraints produce shorter, differently worded answers, which lexical metrics penalise more than semantic ones.

The manuscript is in preparation. I am sixth of ten authors, and contributed the retrieval pipeline and its evaluation.

Recall@10: 0.9640
Precision@10: 0.7527
Recall@5: 0.7129
BERTScore F1: 0.6016

Detector demo

On-device inference on a London street.

A 13-second screen recording from the handset. The overlay shows the model’s telemetry: input downscaled to 640, inference time per frame, p50 and p95 over a rolling window, frame rate and device temperature.

The recording also shows the app’s controls: confidence and NMS thresholds, per-class toggles, and a switch between the pseudo-label-distilled model and its baseline for comparison in the field.

Classes: Cone, barrier, site information board, roadworks sign, other temporary sign
Backend: CPU only, thread count selectable at runtime
Purpose: Local detection reduces the number of frames sent to the server-side vision-language model

On-device inference, recorded 2 September 2026.

Experience

Work experience.

Jul — Sep 2026

Product & Engineering Intern at FYLD

London

Two research workstreams on a field-operations platform: an on-device computer vision model for automated site-compliance checks, and a job duration prediction model. Responsible for the detector, including the dataset, through to architecture sign-off.

Most of the effort went into the dataset. I built it with a model in the loop: automatic labels, mask refinement by a segmentation model, then iterative self-training, with human review of every batch. The final dataset contained 1,400 human-reviewed images and 3,530 boxes from 633 source videos, including 264 true negatives. Ten alternative approaches were tested and rejected along the way, several at a cost of under five dollars each.
Jun — Nov 2025

AI Engineer at Vmor Technology

Sydney

Led development of a real-time speech-to-speech system, built as a reusable module rather than for a single client. Deployed for customer service, automated reception and booking, and clinical note-taking. Peak production load was 1,250 concurrent users.

The pipeline has five stages: a custom PyTorch voice-activity detector, faster-whisper transcription, an LLM with retrieval and an agent layer, open-source speech synthesis, and delivery over WebSocket. It is queue-based and modular, so each stage can be replaced independently. Component choices were driven mainly by cost and data security, so the stack is largely open source. Locally hosted models were also tested, benchmarked and rejected.
Jul — Nov 2025

Research Assistant at University of Technology Sydney

Sydney

Built two systems: a retrieval assistant for pharmaceutical regulatory documents, and a trend-detection pipeline processing more than 10,000 news articles.

The trend pipeline matched articles to known themes by embedding similarity. It identified emerging themes using UMAP and HDBSCAN clustering with near-duplicate filtering, and classified them with supervised models. The regulatory assistant is the subject of a manuscript in preparation. I am sixth of ten authors, and contributed the retrieval pipeline and its evaluation.
Aug 2024 — Jun 2025

Software Engineer Intern at Vmor Technology

Sydney

Backend platform for a member-management product with more than 3,000 users. Two ASP.NET Core APIs over Entity Framework Core and MySQL, covering authentication, payments and membership.

JWT authentication with optional TOTP two-factor authentication, including QR provisioning and expiry handling. Square for card payments, Twilio for one-time passcodes and AWS SNS for notifications, with membership status updated on successful payment. Also profiled and removed N+1 query patterns that were slowing response times.
Feb — May 2024

IT Systems at WollemAI

Sydney

Cloud platform administration at a seed-stage climate analytics startup, including a full migration from Google Workspace to Microsoft 365 for around twenty staff.

The migration had to avoid any email downtime, as the founders were in ongoing client discussions. After the cutover, I diagnosed and fixed an outbound deliverability problem that was sending mail to spam.

Projects

Other projects.

UCL · 2026

Finite difference pricing of European options

A Python solver for European option prices, validated against the Black–Scholes closed-form solution.

Read more Read less

Derived the explicit finite difference scheme for the Black–Scholes PDE and implemented backward time-stepping. Analysed convergence across four grid resolutions, and computed Delta and Gamma directly from the grid.

10.4658 vs 10.4506absolute error 0.0152

Early

Chick embryo developmental staging

Computer vision for classifying the developmental stages of chick embryos, in collaboration with a developmental biology colleague.

Read more Read less

Stages are labelled using the Hamburger–Hamilton staging table. The project applies computer vision to a biology research question.

Outside work

Interests and achievements.

Standing on a ridge with a deep valley and snow-streaked mountain face behind.

On a summit in full winter kit, with a flat sea of cloud below the ridgeline.

Standing on a summit trig point in snow, a sun halo arcing through the sky.

Outdoors: Hiking and winter camping. All photographs on this page are my own, including the one at the top.
Fencing: UCL Men’s 2nd team for two seasons, and team captain for 2026–27.
Chess: Peak online rating of around 1700.
Piano: Currently learning the third movement of Beethoven’s Moonlight Sonata.
Also: Swimming, badminton, drawing and graphic design, Gunpla, laptop hardware upgrades, and cooking.

Vice President, founding team

UCL Actuarial Society. Grew to around fifty members in its first year

Academic Achievement Award

UNSW entry scholarship covering one year of tuition

Dux, ATAR 98.75

Top 1.25% nationally. First in year in Mathematics Extension 2, Chemistry and Physics

Get in touch by email.

Email is the most reliable way to reach me.

email me

Machine learning and AI engineering.

Three projects, and the decisions behind them.

On-device roadworks detection

Problem

Approach

Model selection

Results and validation

Thermal constraints

Data

Stack

On-device inference on a London street.

Work experience.

Product & Engineering Intern at FYLD

AI Engineer at Vmor Technology

Research Assistant at University of Technology Sydney

Software Engineer Intern at Vmor Technology

IT Systems at WollemAI

Other projects.

Finite difference pricing of European options

Chick embryo developmental staging

Interests and achievements.

Get in touch by email.