Project Statement

Face Value is an interactive installation where a digital character mirrors the facial expressions and head movements of its audience in real time. Built with machine vision and facial landmark tracking, the work examines what happens when something as immediate and intuitive as a facial expression is translated into data.

A smile, a frown, or a blink may feel unambiguous to the person making it, but to a computational system, each becomes a set of measurable features mapped onto predefined categories. The avatar responds in ways that are recognisable, yet never quite right. In this subtle mismatch, Face Value makes visible the distance between how we experience ourselves and how a machine learns to interpret us.

By inviting viewers to test the system through their own expressions, the installation turns self-presentation into a live exchange between human and machine. What begins as an attempt to make an avatar mirror us becomes an encounter with the limits of that reflection.

The Brief

The capstone brief was intentionally open-ended: develop an original project representative of the skillset and research built over four years in the program, supported by an artist statement and ongoing peer critique. Beyond "make something that reflects your practice," there were no constraints on medium or scope. We were handed full creative autonomy, and full responsibility for defining the problem worth solving.

Objectives

I'd spent the previous four years demonstrating I could design. For this project, my ambitions quickly became about proving I could build, which meant establishing a constraint at the outset. The interface had to be engineered from the mechanics upward, with an interaction model that was innovative without being too complex.

I did not want to fall back on a conventional WASD scheme simply because it was legible, nor did I want the controls to exist as a functional layer beneath the visual design. The way the player understood the system had to be considered as carefully as the system itself.

The Idea

Gallery installations compete for attention that's in short supply. Attendees can pick from dozens of other projects vying for the same thirty seconds, and almost none of them are willing to sit with a stranger's work long enough to learn a control scheme or wait for a narrative to justify itself. Soon into the quarter, I realised I was effectively designing for zero onboarding in a public, passive-audience context.

It was from this very challenge that an idea was born. I was getting ready one morning, caught my own reflection, and noticed how easy it was to lose several minutes watching my own face move. Nothing was being asked of me in that moment, and yet I couldn't look away. I pictured a project the player could operate using nothing but their own expression, with no controller and no learning curve standing between them and the piece. The logic was almost embarrassingly simple once I had it: every person walking through the exhibit already knew how to make a face.

The Problem

I had never built anything requiring facial tracking, and Unity has no built-in solution for it.

Evaluating Options

My instructors pointed me toward a few paths, such as dedicated facial mocap tools, or rebuilding the project in p5.js/TouchDesigner, an environment I had never used. Both would have meant abandoning either my existing engine fluency or sacrificing my project timeline. Instead, I chose to solve the problem inside the tool I already knew well, which meant finding a way to get real-time facial data into Unity myself.

Finding and Validating the Pipeline

Research led me to MediaPipe, Google's open-source framework for real-time, on-device ML pipelines, and from there to a community-built Unity plugin. I validated the pipeline in stages before committing:

• Imported the plugin and ran its sample scene to confirm whether webcam input and face-mesh tracking worked at all.
• Confirmed the plugin's Face Landmarker tools could generate actual blendshape values by modifying the sample script to print blendshape scores to console.
• Only once both were confirmed did I move on to using that data for anything.

This staged validation made sure that every subsequent hour of work was built on a foundation I'd already proven, rather than discovering a dead end deep into production.

Building the Input-to-Animation Bridge

MediaPipe's blendshape output is normalized (0–1); conveniently, the same scale Blender's shape keys use. That meant the cleanest path to an actual character response was a direct mapping: MediaPipe blendshape → Unity script → Blender shape key.

I built and tested this on a simple placeholder head — "Thrawn," after the Star Wars character — with four shape keys, to confirm the technical pipeline worked before spending time on character design.

After importing Thrawn into Unity, I wrote a bridge script that exposed each shape key to the Unity inspector, then mapped MediaPipe's output values to those fields in the tracking callback: a smile from the player increases mouthSmileLeft and mouthSmileRight (0–1) → the driver stores it → converts it to a 0–100 shape key value → the character animates using its SmileLeft and SmileRight sliders.

Once Thrawn animated correctly from live webcam input, the project's fundamental feasibility was no longer in question. That milestone landed at the project's midterm review.

Scaling Up and the Problems That Came With It

Moving from a 4-key test head to an 11-key final character (blinks, smiles, frowns, brow motion, jaw) surfaced problems that hadn't existed at small scale:

• Raw mapping looked exaggerated and uncanny — MediaPipe's blendshape values are tuned for maximum expressiveness, and a 1:1 mapping made the character look like it was overacting. Addressed with intensity multipliers, clamped maximums, smoothing, and dead zones hand-tuned per shape key.
• Some tracking data was simply unreliable — frown detection barely registered even at maximum effort. Compensated with a targeted amplification of that response curve specifically, since over-amplifying everything would have reintroduced the exaggeration problem.

Process & Iteration

With the core interaction proven and the concept honed, remaining work moved quickly:

• Head/neck rig driven by tracked head movement, with heavy smoothing and dampening, which reduced uncanny jitters and made viewers recognize themselves as the clear source of the character's motion.
• Three full characters and environments (Gem, Pinkie, Aurelian respectively), each built on the same 11 shape keys so the same driver script and mapping logic could be reused across all of them.
• A timer-based scene cycler instead of a menu or intro screen — a choice to keep zero onboarding as a hard constraint through to the final build.
• Cross-demographic playtesting specifically to check for tracking bias across different faces, since MediaPipe's accuracy wasn't guaranteed to be uniform.

Deployment Engineering

Getting a working prototype to survive 24/7 operation for two weeks unattended in a gallery surfaced an entirely separate category of problem from the design work:

• Viewers stepping out of frame left the character frozen on their last expression. Fixed by scripting a timeout that zeroes all blendshapes when no face is tracked for a set duration, resetting to neutral for the next viewer.
• The project was initially developed on a Mac with a built-in camera, but the exhibition used an external USB webcam on Windows. I created a camera fallback system ("use the preferred device if present, otherwise fall back to the first available camera") to guarantee the build wouldn't fail to launch on unfamiliar hardware.
• The Windows build failed to initialize the webcam at all despite correct permissions. Diagnosed via player.log that MediaPipe's LocalResourceManager only functions inside the Unity Editor, not in a built executable. Resolved by switching to StreamingAssetsResourceManager and manually reorganizing the plugin's resource files into StreamingAssets so they'd be accessible at runtime.

Validated all fixes with a full 24-hour unattended stress test before committing to the install.

Physical Install

With the software validated, I installed the piece independently in the gallery: wall-mounted the display via VESA mount, fabricated and painted a custom ventilated enclosure for the mini PC (no floor or table space was available), and handled cable management for a clean public-facing presentation.

Tools & Pipeline

Result

Face Value opened on June 4, 2026, and ran continuously through de-installation on June 15. During gallery-sitting shifts, I observed that nearly every passing viewer stopped to engage far longer than a typical gallery glance. They never once needed an explanation of the controls. The project achieved its core design goal: an installation that captured passive audiences immediately, using an interface as natural as making a face.

Project Statement

An investigation into the narrative potential of 3D space. As my introduction to 3D modelling, the project provided an opportunity to consider what an environment can contribute beyond its conventional role as setting: how spatial relationships, atmosphere, materiality, and the boundaries of a physical space can communicate character and emotion in their own right. The room and cave consequently function as extensions of the psychological state they contain.

Solitude can offer a sense of autonomy and refuge, but the boundary between chosen solitude and self-isolation is not always easily maintained. Conceptually, this short film considers the psychological tension that comes with the awareness that a familiar form of retreat has become confining, coupled with the inertia that makes leaving it feel almost impossible.

Project Statement

A digital portrait study exploring the relationship between visual perception and personal interpretation. The project examines how viewers construct narratives around unfamiliar faces, projecting an assumed identity onto a subject based solely on the limited context available to them.

Presented through orbiting lights across a static 3D head, the piece uses lighting as a tool for revealing how perception changes even when the subject itself remains unchanged. As shadows move across the face, subtle variations in expression and mood might emerge based on the viewer’s own interpretation.

Project Statement

Alienation is an interactive exploration of what it means to navigate a world where belonging depends on understanding rules that were never written for you. Set on a hostile alien planet, the game follows a human disguised as an alien who must infiltrate an unfamiliar society by taking on the role of a chef. To survive, the player must perform everyday tasks while concealing the fact that they do not truly understand the world around them.

The game translates the experience of being an outsider into its core mechanics. In the opening grocery store sequence, players navigate a disorienting maze-like environment in search of ingredients, mirroring the confusion of entering a space where familiar structures no longer apply. The cooking segment further reinforces this disconnect by presenting recipes written in an unreadable alien language, forcing players to react to unfamiliar information through experimentation and intuition alone.

Project Brief

This project simulated the structure of a professional design studio commissioned by a client to develop their brand. Working in teams of four to five, we established an original organisation, defining its purpose, audience, positioning, and strategy before translating those decisions into a complete design system.

Over the course of ten weeks, we developed the system's core elements—i.e. its logo, colour scheme, typography, motion, and material language—and applied them across multiple touchpoints including print, digital, and environmental design. The final stage was to consolidate the system into a set of brand guidelines, presented as though we were proposing the identity to a client.

Rather than simply producing a collection of assets, the project required us to build an organization with a clear point of view and a coherent internal logic, then demonstrate how that identity could operate consistently across different contexts.