I'm an AI Engineer who builds full-stack, AI-powered products end to end, from tool-calling agents and RAG pipelines to shipped, deployed apps. My background training deep learning models on medical imaging data gives me a research-grade rigour about evaluation and edge cases that a lot of AI engineering work skips.
Medical Imaging Science Univ. of Manchester, Merit
BEng
Biomedical Engineering Shantou University, Top 30%
Who I Am
Building AI end to end from model to product
I hold a BEng in Biomedical Engineering from Shantou University and an MSc in Medical Imaging Science from the University of Manchester, where I graduated with Merit, a background that gave me a solid foundation in deep learning before I moved into building full-stack AI products. I'm a graduate visa holder, so no sponsorship is needed.
What I enjoy most is taking a problem and working through it systematically end to end: designing a RAG pipeline, wiring up an LLM provider, shipping a working app, or further back, writing a segmentation pipeline and reviewing hundreds of landmark annotations. Same mindset, different stack.
The domain background behind that: CT and MRI processing, image segmentation, statistical shape modelling and 3D reconstruction, plus the research habits that come with a lab, experimental design and writing up what you find.
AI / LLM Engineering
Tool-Calling AgentsAgent EvaluationVercel AI SDKRAGLLM IntegrationStructured OutputEmbeddingsPrompt Engineering
It started in a microfluidics lab in Shantou, led me to imaging research in Manchester, and from there into building AI products end to end. Here is how it unfolded.
01.2025 – Present
Independent AI / ML Developer
Self-employed
Shipping AI applications end to end for small businesses, from the model to a deployed product.
Designing and shipping AI applications end to end for small businesses, taking each one from the model through to a deployed product. Most of my recent work has been on tool-calling agents and on how you actually measure whether one is any good.
Built RadReport Agent, a tool-calling clinical imaging agent that chooses its own tools across a chest X-ray classifier, lung segmentation, cardiothoracic ratio geometry, report retrieval and PubMed, then answers with every claim checked against the raw tool output. Scored 97.7% tool selection and 96% groundedness on a 43-case gold set.
Built PantryChef, a Next.js/TypeScript recipe app with RAG-based chat (embedding index + cosine retrieval), integrating three external APIs and a Nutri-Score health grading layer; deployed to production with automated tests.
Built Chouette, a French-shadowing app pairing synced YouTube transcript playback with bring-your-own-key AI conversation practice, backed by Auth.js, Drizzle ORM and Postgres.
Built a full-stack medical image analysis platform (React, ASP.NET Core, FastAPI) running pretrained MONAI segmentation models on uploaded CT scans, with JWT auth and an interactive 3D viewer.
Shipped NoteWidget, a native macOS always-on-top sticky-note widget in SwiftUI/AppKit with a global hotkey, tagging and local SwiftData storage.
Four months hands-on with cell culture, single-cell library prep and PDMS chip fabrication.
An intensive four-month immersion at one of China's leading genomics institutes. Worked hands-on across four areas: cell culture and droplet generation for microfluidic experiments; C4 scRNA-seq V2 single-cell library preparation; fabrication and calibration of PDMS microfluidic chips; and barcode modification of magnetic beads using multi-channel pipettes. Coordinated daily with three other lab teams (wet-lab, chip fabrication and QC) to hand off samples and align on protocol changes, which meant a lot of cross-team communication to keep experiments from stalling. Gained fluency with electron microscopes, plasma cleaners, sorter machines and vacuum equipment. Learned to read experimental proposals independently and iterate when experiments failed.
A year spent diving deep into CT and MRI analysis, imaging mathematics and computational methods. My dissertation focused on 3D vertebral shape modelling from CT scans, where I worked on landmark annotation and compared different model configurations for anatomical localisation.
CT ImagingMRIShape ModellingPython
03.2022 – 03.2023
Research Assistant, Microfluidics Laboratory
Shantou University
My first real taste of lab research, working on a microfluidic chip designed to detect lead ions in water. I ran the chip testing, managed the fluid control experiments and helped analyse performance data. We got reliable detection down to 0.1 ppm, which felt like a genuine win.
MicrofluidicsLab ResearchData Analysis
09.2019 – 06.2023
BEng (Hons) Biomedical Engineering
Shantou University, China · GPA 3.4/5.0 · Ranked Top 30%
Four years building the foundations in medical imaging physics, computational simulation and bioinformatics. Graduated with an average of 84/100 and a real curiosity for how engineering and medicine could work together.
Biomedical EngineeringMATLABMedical Physics
Sep – Dec 2024 · BGI Research
Internship in Pictures
Four months at the National Gene Bank in Shenzhen, surrounded by cutting-edge genomics labs, experimental animals, and microfluidic equipment. A glimpse of the environment and daily lab work.
Campus Life
Ostrich at the BGI campus grounds
Campus Life
Sika Deer roaming the National Gene Bank
Research Environment
Vertical plant growth unit in the elevator lobby
Lab Work
Examining a microfluidic chip under the microscope
Lab Work
Modifying magnetic beads using a multi-channel pipette
Lab Work
Operating the microfluidic sorting setup
What I've Built
Project Highlights
The five projects that best show where I am now: a tool-calling clinical agent and full-stack AI builds, alongside medical imaging and deep learning work. Five more from my biomedical engineering background and other builds are one click away.
Give it a chest X-ray and a question, and it works out which tools to call.
Give it a chest X-ray and a question, and the agent works out which tools to call. It has six: a DenseNet-121 classifier, PSPNet lung and heart segmentation, deterministic cardiothoracic ratio geometry, BM25 report retrieval, exact case lookup and PubMed search. What comes back is a structured, cited answer where every clinical statement traces to the tool call that produced it. It is a research prototype built on the public, de-identified Indiana University chest X-ray collection, and it is not a medical device.
Built the agent loop over six tools with a handle-passing design: segmentation writes masks to disk and returns a handle the CTR tool consumes, so large arrays never enter the model’s context
Enforced grounding with Pydantic-validated structured output plus a verifier that checks every quote against raw tool output and rejects fabricated ones, which took groundedness from 52.9% to 96.0%
Wrote a 43-case gold-set evaluation harness with per-category scoring: 100% convergence, 97.7% tool-selection accuracy, $0.00102 and 11.7s median per query
Shipped a Streamlit UI with a live trace panel showing every tool call, its arguments, latency and returned JSON, plus one-click adversarial cases for missing data, out-of-scope requests and fabrication bait
Containerised the stack and verified it properly: 161 tests pass inside the image with --network none, which surfaced three defects the host suite could not
Deployed to a ~1 GB free tier by serving precomputed real model outputs behind a demo flag, cutting peak memory from 1,816 MB to 71 MB while retrieval and the agent loop stay fully live
Upload a CT scan, run AI segmentation on it, and explore the result in 3D in the browser.
A full-stack demo platform for uploading CT scans (NIfTI format), running an AI segmentation model on them, and viewing the results in an interactive 3D viewer in the browser. An educational/portfolio project, clearly scoped as research-grade and not validated for clinical use.
Built a three-tier stack: React/Vite frontend, ASP.NET Core API for auth and orchestration, Python FastAPI service for inference, with PostgreSQL for metadata via EF Core
Integrated two pretrained MONAI Model Zoo bundles, a 3D UNet for spleen and a SegResNet trained on 104 structures for liver, kidneys, gallbladder, stomach, pancreas and bladder
Implemented JWT-based auth, NIfTI upload, and volume visualisation with segmentation overlays via Niivue
Surfaced segmentation stats (voxel count, estimated volume in mL, inference time, model used) and containerised the whole stack with Docker Compose
Turns whatever is in your kitchen into matching recipes, a health grade and a grounded Q&A assistant.
Built a full-stack app that turns whatever's in your kitchen into matching recipes and cooking videos, complete with a Nutri-Score-style health grade and a RAG-powered Q&A assistant grounded in a hand-curated meal-prep and nutrition knowledge base.
Built a retrieval-augmented generation pipeline (embeddings + cosine similarity search) grounding an LLM chat in a custom nutrition knowledge base, with cited sources
Implemented an explainable A–E health score by adapting the Nutri-Score formula to each recipe's nutrition facts
Integrated Spoonacular, YouTube Data API and Open Food Facts, with aggressive caching so the app degrades gracefully under tight free-tier quotas
Designed a pluggable LLM provider layer (Groq / Anthropic / OpenAI) so a public demo link never runs up someone else's API bill
Locating vertebrae in CT scans with landmark annotation and statistical 3D shape modelling.
Built a pipeline to localise vertebrae in CT scans using landmark annotation and statistical 3D shape modelling. A detail-heavy project that taught me a lot about the gap between a working model and a reproducible one.
Reviewed and refined 345 vertebral landmark annotations to improve training data quality
Compared single-stage vs two-stage configurations on localisation accuracy and runtime
Wrote up the full workflow so others could actually reproduce it
Three segmentation models compared on 200 head CT volumes, tuned for small lesions.
Developed and evaluated three segmentation models — a 3D U-Net baseline plus improved 3D and 2D variants — on 200 non-contrast head CT volumes with voxel-level expert annotations, with preprocessing targeted at small-lesion performance.
Built brain-ROI extraction, HU windowing, median-spacing resampling and elastic/affine augmentation into the preprocessing pipeline
Raised mean Dice from 0.60 to 0.66 (3D) and 0.68 (2D), and cut Hausdorff Distance from 50.0 to 29.4
Configured the CUDA 11.0 / PyTorch 1.7 environment from scratch on a remote Linux workstation (SSH, single RTX 3080) within a 10 GB VRAM budget
Shadow French with leveled YouTube clips, then sync the words you saved straight into Anki.
Built a French-learning app centred on shadowing: leveled YouTube clips with synced transcript playback, tap-to-save vocabulary, and an AI conversation partner, all synced back into the learner's own Anki deck.
Built a synced shadowing player with line-by-line transcript playback and record-yourself comparison across A1–B2 leveled clips
Added bring-your-own-key AI conversation practice that flags mistakes inline and files them into a personal error notebook
Wired one-click vocabulary sync to Anki via AnkiConnect, plus GitHub/Google auth via Auth.js for per-user data
Built on Next.js, Drizzle ORM and Postgres (Neon), deployed on Vercel
Next.jsTypeScriptDrizzle ORMPostgreSQLAuth.jsVercel AI SDK
May 2022 – May 2023 · Shantou University, Department of Biomedical Engineering
A low-cost chip that reads lead ion levels in water without a lab instrument.
Designed and prototyped a point-of-care testing (POCT) microfluidic chip capable of quantitatively detecting lead ion (Pb²⁺) concentrations in natural water samples. The chip provides a low-cost, portable alternative to traditional laboratory instruments such as atomic absorption spectrometry, achieving a detection limit of 0.44 ng/mL and producing results readable by the naked eye or a smartphone camera.
Chip architecture: self-driven capillary pump with PMPs capture channel and hydrophilic pad
Laser-cut hydrophilic adhesive tape fabricated for the chip's wetting layer
Chip yield rate before fabrication process improvements
Designed the chip architecture in Fusion 360 and SolidWorks, including the dual-layer "hydrophilic tape + OSTE micropillar" structure to improve wettability and extend shelf life
Proposed use of OSTE (off-stoichiometry thiol-ene polymer) as the chip material for its rapid UV-curing, long-term stability and commercial scale-up potential
Developed a PDMS-to-OSTE two-stage moulding process to prevent damage to fine photomask features during demoulding
Achieved a detection limit of 0.44 ng/mL, substantially lower than conventional FAAS methods (typically 1-80 ng/mL), with results quantifiable by naked eye or smartphone
Analysed sensitivity with MATLAB and ImageJ; created project timeline and Gantt chart to coordinate a 3-person interdisciplinary team across 12 months
Sep 2021 – Jan 2022 · Shantou University, Department of Biomedical Engineering
Senses harmful gas and humidity, then opens your window over Bluetooth and texts you.
Designed and built a smart indoor air quality monitoring and control system as a team using the STM32F407ZG ARM Cortex-M4 MCU. The system detects harmful gas and CO concentrations, regulates humidity automatically, and allows the user to remotely control a window motor via Bluetooth, with real-time SMS alerts delivered by a GSM module.
Programmed STM32F407ZG in STM CUBE IDE using USART, ADC and GPIO protocols to integrate sensors, Bluetooth, and GSM modules via AT commands
Implemented 3-level air quality LED indicator (green / yellow / red) using dual sensors for CO and VOC harmful gas
Enabled remote window actuation via Bluetooth (stepper motor clockwise/counterclockwise) and automated humidity control via atomiser (triggers below 65% RH)
Delivered real-time air quality SMS alerts to the user's phone via GSM module; temperature and humidity displayed live on LCD
STM32Embedded CUSART / ADC / GPIOBluetoothGSMSensor IntegrationSTM CUBE IDE
Demo: STM32-based smart air purifier, real-time air quality monitoring and automated control
Undergraduate Group Project
🔬
Nose Septum Rongeur: Ergonomic Redesign
Sep 2021 – Jan 2022 · Shantou University, Biomedical Engineering
An adjustable opening-angle mechanism so the tool matches the patient, not the other way round.
Redesigned a nose septum rongeur to address a key ergonomic limitation: the maximum release range could not be adjusted to match different patient anatomies, requiring surgeons to exercise extra caution to prevent accidental injury. Our team added an adjustable maximum-opening-angle mechanism using a grooved rod and circular card buckle, a structure brainstormed entirely by the team without referencing existing concepts.
Designed a rod-with-grooves and circular card buckle mechanism to limit the maximum scissor-head opening angle, protecting patients and reducing surgeon hand fatigue
Built a full CAD assembly and 3D-printed a functional prototype for side-by-side comparison with the original surgical tool
Validated design via FEA (AISI 304 stainless steel, 30–65 N load range): max rod displacement of 0.628 mm under 65 N, confirming structural integrity of the grooves
Performed mesh convergence study demonstrating minimal sensitivity of displacement results to mesh refinement
Medical Device DesignFEA / SimulationCAD3D PrintingSurgical InstrumentsStainless Steel
Prototype demonstration: adjustable opening-angle mechanism in action
Undergraduate Group Project
🥽
Mixed Reality Guidance for Breast Cancer Tumour Removal
HoloLens overlays that track tumour movement live to keep resection margins under 1 cm.
Proposed a Mixed Reality–assisted surgical system to improve the accuracy of breast tumour resection. Using a Microsoft HoloLens headset, pre-operative imaging data is processed into real-time 3D overlays that track tumour movement during surgery, reducing normal-tissue excision error from the current 1–2 cm margin to under 1 cm.
Designed a full surgical workflow: pre-op data collection via medical imaging → 3D tumour model generation with smoothing treatment → real-time overlay projected onto the surgical field via HoloLens
Addressed motion artefacts by implementing a feedback loop: glasses detect tumour displacement and transmit updated positional data to the computer, which re-renders and re-projects the image in real time
Improved projected AUC from 0.65 (current standard) to 0.90 for sub-1 cm normal-tissue resection accuracy
Conducted QALY analysis: treatment QALY increases from 0.85 to 0.95; differential benefit of 350,000 CNY including societal gains; incremental device cost 400–1,100 CNY per patient
Assessed patentability (novel real-time Mixed Reality surgical guidance), ethical compliance (beneficence, autonomy, green engineering), and social impact (reduced patient–doctor conflict via shared pre-op visualisation)
Mixed RealityHoloLensSurgical Guidance3D Image ProcessingQALY AnalysisMedical ImagingBreast Cancer
Beyond the Lab
My Interests
What I get up to when I'm away from the keyboard.
🧗
Rock Climbing
Climbing is my favourite way to switch off. There is something I love about how it forces you to be fully present, physically and mentally. Every route is a puzzle and I am hooked.
Last year I travelled solo across several countries and wrote up guides to help others do the same. Going alone sounds intimidating but it is honestly the best way to actually experience a place.
I am looking for a role where I can contribute from day one and keep growing. If you're building AI-powered products and think we might be a good fit, I would love to hear from you.
Roles I'm Open To
🤖
AI Engineer / ML Engineer
Building and shipping AI-powered products end to end, from RAG pipelines and LLM integration to production deployment.
💻
Full-Stack AI Developer
Next.js, TypeScript and the Vercel AI SDK, turning an idea into a deployed, working product fast.
🧠
Applied AI, Healthcare & Life Sciences
Bringing deep learning experience from medical imaging (segmentation, 3D shape modelling) into AI systems that matter in healthcare.
Available Now · UK-based · No Sponsorship Required
