skip to content
Felicia photographing the street with a compact camera, reflected in a car wing mirror in San Francisco's Chinatown.

hi, i'm felicia.

cs + physics @ stanford. student of life above all.

research interestscomputer visionspatial intelligenceworld models

some things i've worked on

01

MAESTRO

seen · framessaid · speech → semantic chunksheard · non-speech audiodrum hit · punchlineunified representationVSVASVagentzoom inobjectsactionsfacesverdict+ explanationschematic of the pipeline · not experimental data

Video models tend to overlook sound. MAESTRO fuses what is seen, said and heard, then reasons over the moments that matter, outperforming GPT-4V on hateful-video detection with no dataset-specific training.

93% F1 · Regeneron ISEF 2nd Place

Recognised by Singapore's Minister of Defence

read the research page →
02

Meta-Robots

demo · a single episode

An agent studies why a navigation system fails, proposes a fix, implements it and measures the effect, lifting HM3D ObjectNav success from 65.4% to 68.9% over the ASCENT baseline.

68.9% HM3D ObjectNav success · +3.5 pp over baseline

read the research page →
03

SINDEX

input · claim + imagephoto shared with the claimevidence · retrieved livekeywordsnews search25 closest articlestop 3 articles · 2 sentences eachthree checkstruthfulnessclaim + image + evidencesatireis it exaggerated for effect?out of contextdoes the image belong here?misinformation · satire · out-of-contextverdict+ explanation citingthe evidence it foundschematic of the pipeline · not experimental data

Fact-checking that shows its work: it retrieves live news evidence, checks whether a claim and its image really fit together, and tells misinformation, satire and out-of-context images apart, built for Singapore's context.

95.2% F1 · IRC-SET Best Paper

Recognised by Singapore's Minister of Defence

read the research page →

currently building...

more projects

    • Co-Founder & Chief Technology Officer (CTO), 2026 – present
    • Spearheaded the technical architecture and end-to-end development of Zing
    • Engineered a multimodal AI content-generation pipeline integrating Claude Sonnet and ElevenLabs, with production infrastructure on Supabase and Railway
    • Led product engineering from concept to public launch on the Apple App Store and Google Play, securing institutional contracts with 7 GP tuition centres and government schools in Singapore
    • Tech stack: Claude Sonnet, ElevenLabs API, Supabase, Railway
    • Deployed at three farms in Hainan
    • 27% efficiency improvement
    • International Champion, Young Inventors Challenge
    • Developed a web AR experience used by 1000+ participants to resolve queue bottlenecks
    • Engineered a system to track 30k of event logistics, replacing manual processes
    • Designed the UI and built location-based challenges
    • Released on the App Store and presented to Apple's Southeast Asia Head of Development
    • Built through the Swift Accelerator
    • Next.js + Firebase
    • ~2,400 students served, ~400 monthly active users

physics research

Modelling and experiment under imperfect physical models: the same simulation-to-reality problem that embodied AI runs into.

    • Investigated non-contact rotational energy transfer between magnetically coupled systems, studying how electromagnetic interactions govern angular motion and synchronization
    • Developed numerical models in Wolfram Mathematica to simulate coupled rotational dynamics, incorporating magnetic torque, friction, and moment of inertia
    • Compared simulated trajectories against experimental observations, investigating discrepancies between theoretical predictions and physical system behavior
    • Investigated the nonlinear behavior of fluid jets interacting with inclined porous surfaces, examining how flow conditions and surface geometry influence jet deflection
    • Applied computational fluid dynamics (CFD) and experimental analysis to study jet trajectories, fluid–surface interactions, and the underlying mechanisms governing refraction
    • Evaluated theoretical models against experimental observations and defended findings through scientific presentations and technical debate

stem empowerment work

    • 210+ girls reached across 10 schools, with monthly workshops
    • 50-hour AI/web incubator: 60 participants building for social causes
    • Science Castle partnership
    • 11 participants presented research to an international audience from five countries
    • visit →
    • Secured $20K in grant funding to launch the programme
    • Co-developed a scalable curriculum reaching 800+ students and trained 50 educators, enabling long-term sustainability and expansion
    • Co-planned and organised a youth volunteering fair for 327 attendees