Computer vision
Gesture recognition
One reference image, no training. The recogniser compares hand graphs by their Laplacian spectra, live from a webcam.
First written in 2020, then extended to full-body tracking for a 2022 installation.
Lefinno Kwok: engineering and craft
01 / Engineering
Engineering
Management, design, communication and technology. Broad by default, deep where it matters.
CurrentlyAnalyst, DeloitteEnterprise platform, 2026 to present
Enter engineering02 / Craft
Craft
Games, tools and objects that are fun to use, and useful when it counts.
In productionEccentricitySteam, Chefs Hat Studio
Enter craftEngineering / 01
Broad by default. Deep where it matters.
I am a generalist with a technology specialization. I move between management, finance, communication, art and design, and engineering. AI and ML are where I go deepest technically, not the whole story. Most of what I do is take an idea from conversation to something real.
01
Selected work
2025 to 202702
Experience
2021 to now2026 to now
DeloitteAnalystEnterprise AI platform engagement: identity, data access, alerting and offline deployment. Client withheld.2025 to 2026
HKUCOMP4801 capstoneProbabilistic decision support system, built and evaluated solo. Supervisor: Professor Liang Yingyu.2024 to 2025
DeloitteDigital Foundry InternDeployed a digital practice for a luxury retail client and pitched an IoT measurement system.2024
HKSTPUniversity Collaborations InternBuilt a small CRM that lifted query throughput from about 70 to about 400 per person per day.2021 to 2022
AI MnemonicElectronics then software engineerThree embedded sub-projects; video transmission improved from 0 to 16 FPS.03
Capabilities
Range and depthGeneralist breadth with specialist depth. I am most useful at the seam between disciplines: scoping the idea, building the system, and translating it for the people who have to run it.
01General / range
Soft skills and interests
- Management
- Finance
- Leadership
- Communication
- Art and design
- History
- Philosophy
02Technology
The buildable range
- Platform engineering
- Product
- Software
- Electronics
- Hardware at transistor level
- Mechanical and chassis design
03Specialization
The technical depth
- AI and ML
- Systems
RangeThe wide layer is how I read people, money, taste and context: management and finance, leadership and communication, art and design, history and philosophy.
TechnologyNarrower and longer: the technical range that makes an idea buildable. Platform and product, software, electronics, hardware down to transistor level, mechanical and chassis design.
DepthThe narrow end is the specialization: AI and ML, and the systems that make them operable.
01Range
Management and finance
Scoping, proposals, budgets, effort estimation and vendor coordination. Comfortable with the commercial side, not just the build.
02Range
Leadership and communication
Coordinating small teams and cross-organisation stakeholders, writing documentation and translating between technical and non-technical people.
03Range
Art and design
Visual identity, interaction, 3D and game art. Blender, Inkscape, FreeCAD and front-end craft; I care how the thing feels.
04Range
Product and software
TypeScript and React, Go, Rust, Java and Python. Frontends, APIs, data access and integration; typed boundaries over glue.
05Specialization
AI and ML
LLM orchestration, evaluation design, Bayesian decision models, computer vision, uncertainty and provenance. A key part of the technology specialization.
06Depth
Systems and infrastructure
Linux, Docker, Kubernetes, networking, offline deployment, high availability, observability and backups.
07Range
Hardware and interaction
KiCad boards, ESP32 firmware, FreeCAD and 3D printing, game engines and interactive installations.
08Range
Research and writing
Literature-grounded design, baseline evaluation, failure analysis and technical writing. Publishing limitations, not just results.
04
Current
Analyst, Deloitte
Integrating an enterprise AI agent platform for a large client: identity, data access, alerting and offline porting. The client, the vendor and the systems stay out of this site by design; the CV carries the detail.
SinceSeptember 2026
FocusEnterprise platform integration
StackGo, Rust, Java, Kubernetes
Based inHong Kong
Craft / 02
Fun to use. Useful when it counts.
Games, tools and physical objects, judged on whether someone enjoys using them and whether they help. Some of it is commercial, some of it is a jam, all of it is built for someone other than me.
01
In production
2025 to now2026
Eccentricity
A steampunk engineering roguelike about building chain-reaction power grids. It started as a Ludum Dare entry and became a commercial release with a data pipeline, tests and a repeatable build process.
_
02
Demos
Run them hereFive working systems from both sides: a gesture installation, a transistor computer, a training playground, a CNN inspector and a language-to-diagram tool.
Simulation
8-bit CPU simulator
A computer built from discrete transistor logic. Step the microcode, run a program, edit assembly and watch the bus.
The machine was built in 2020 from discrete logic; the browser simulator is the 2026 rebuild.
Machine learning
ML playground
Draw a digit and classify it, or train a small classifier and watch the decision boundary move as it learns.
A teaching tool from the capstone era: small models you can train and inspect in the browser.
Machine learning
CNN visualiser
A convolutional network reading digits in three dimensions: layers, weights, activations and the decision.
The visualisation layer from the ML playground, separated so the network can be read in three dimensions.
How it worksOpenClose
- 01
Input
The current diagram and your message. The demo keeps conversation state and supports class, sequence, activity, use case, component, state and object diagrams.
- 02
Prompt
A fixed system prompt sets the role and the output contract. The current PlantUML source is included so an edit is targeted, not a full regeneration.
- 03
Model
Any OpenAI-compatible endpoint with your own key, at temperature 0.2. The shared proxy is gone, so no model call happens without your key.
- 04
Parse
The response must start with CODE or NEED_CLARIFICATION. The parser extracts the fenced block, strips stray plantuml lines and repairs missing @startuml / @enduml delimiters.
- 05
Encode
UTF-8, raw zlib deflate, Base64, then PlantUML's custom URL alphabet. This runs in the browser with pako; no source leaves for encoding.
- 06
Render
PlantUML's public server receives only the encoded source and returns SVG. The image, PNG export and editor link all come from that one artifact.
Language models
UML generator
Describe a system in natural language and iterate with a model until the class diagram matches what you meant.
Built in 2025, when iterative LLM diagramming was still novel. In 2026 the interaction matters more than the generation. The shared proxy is retired; bring your own API key.
04
Chronology
2018 to 20262018
First gameSpace Saver, Ludum Dare, high school.2020
Transistor computerDiscrete logic, custom instruction set.2022
Retro handheldCustom PCB, firmware, enclosure.2024
Intern giftESP32 LED medallion for the HKSTP cohort.2025
EccentricityJam build to commercial production.2026
Studio workSteam page, Binmin, contract development.Practice / 03
An idea of how I work.
Not a fixed process. A rough shape that tends to appear whether the work is a service, a board or a model. The drawings below are artifacts from that practice, not a catalogue of the only shapes I work in.
01
Working notes
Usually, not alwaysListen
Usually starts with talking to everyone who touches the problem: users, operators, buyers, the people who maintain it. Separating the request from the need is most of the work.
Imagine and estimate
Then research, references, raw material and resources, a few rough alternatives, and a guess at effort and risk before committing to one.
Structure
At some point the idea has to take a technical shape: boundaries, data, interfaces, failure modes and a build order.
Build
Building tends to happen in slices. Integrate early, put working pieces in front of the people who asked, and let feedback change the plan.
Prove and hand over
Eventually it has to survive real inputs and real operators. Documentation, a runbook, and enough context that someone else can run it.
A sketch, not a procedure. The order changes, the loops are real, and listening and building keep feeding each other. The two artifacts below are examples from that practice, not a taxonomy of the work I can do.
02
Artifacts
Two examplesArtifact · Service shape
Service architecture
The architecture I reach for: typed clients, a gateway, service boundaries and a build path, with each module tied to a language.
Output of Model and Verify: boundaries, interfaces and the build path before the first service is written.
Artifact · Hardware path
Hardware pipeline
From schematic to firmware to assembly, with the tools I actually use at each stage and where the branches rejoin.
Output of Model and Operate: the actual route from schematic and board to firmware and assembly.
About / 04
A generalist with a technology specialization.
I am a builder in Hong Kong. My range runs from management and finance to communication, art and design, with technology as the specialization I use to bring projects to life. AI and ML are where that specialization goes deepest, not the whole identity.
I studied computer science at HKU, worked across an embedded startup, a science park and a consultancy, and now integrate an enterprise AI platform. I care about the unglamorous middle: integration, tests, release processes, and the parts that decide whether a system survives contact with reality.
Based inHong Kong
LanguagesCantonese, Mandarin, English
EducationBEng Computer Science, HKU
CurrentlyAnalyst, Deloitte
StudioChefs Hat Studio
01
Trajectory
2015 to now2015 to 2020
Making
Microcontrollers, a game jam entry, and a computer from discrete transistors.
2021 to 2024
Working
Embedded video, automation at HKSTP, then enterprise delivery at Deloitte.
2025
Research
A solo capstone on probabilistic decision support, evaluated against five language models.
2026
Now
Analyst at Deloitte, studio owner at Chefs Hat, and an event platform for 300 to 400 jammers.