Hi, I'm Nathan Cheng 程一

Scroll
§ 01

Currently

Final year of my Bachelor of Psychology at HKU. Co-founder at SmartCited, and working with Anne Collins' lab at UC Berkeley on neural-network parameter estimation and model identification.

Nathan Cheng
Hong Kong · 2025
§ 02

Selected AI product work

AI visibility actions · Smart Cited Inc.
SmartCited AI-search citation intelligence interface
Selected 01Co-founder · 2026

SmartCited

Co-founded an AI-visibility action system for local businesses: it asks 10 leakage-checked discovery questions through OpenAI and Gemini, preserves the citation evidence, then prepares customer-authorized website, Yelp and directory actions. Started at OpenAI Build Week, incorporated as Smart Cited Inc. in August 2026.

630-answer benchmarkOpenAI + GeminiAuthorized actions
Live product · Public view
PivotBridge AI career agent interface
Selected 02Berkeley SkyDeck · 2026

PivotBridge.AI

Evaluated a live AI career agent across onboarding, interview, resume and decision workflows — turning trust gaps and UX friction into implementation-ready product recommendations.

Product evaluationAI output qualityShipped improvements
Visit product
Wellbeing AI · Psychology research
Welldify wellbeing platform homepage
Selected 03Hong Kong · 2025

Welldify

Evaluated an LLM trained on 66,000 labeled sentences for journal-based wellbeing scoring against PERMA and DASS frameworks, and built a semi-automated DeepSeek and generative-image workflow that shipped 30+ psychoeducation carousel sets.

LLM evaluationPERMA + DASS30+ carousel sets
Visit product

VStandby Studio · 24-hour BETA AI hackathon build: a real-time AI companion that detects livestream highlights, follows caption context and reacts through a selectable persona.

Open demo ↗

Everyone At Once · Independent 0→1 product: twelve situational questions become a mix of fictional characters through LLM simulation and Bayesian inference.

Try it live ↗
§ 03

Research experience

'26UC Berkeley
w/ A. Collins
CCN Lab

Built the simulation-based inference pipeline (TensorFlow GRU) that recovers reinforcement-learning-with-forgetting parameters and tells that model apart from a working-memory model in simulated behavior — at 27,000 agents per model the classifier reaches 0.92 accuracy against a 0.975 likelihood benchmark. Second author on a tutorial manuscript in preparation with Ti-Fen Pan and Anne Collins.

simulation-based-inferenceneural-netscognitive-modelstensorflow
'25HKU
w/ F. Cao
NeuroLeaD Lab

Developmental Trajectories of Brain Functional Connectivity in Autism Spectrum Disorder — third author, invited revision at PNAS, 2026. Deployed neuroimaging pipelines on HKU's HPC (a first for the lab) to process ABIDE-I/II and ADHD-200, and ran a preliminary brain-age analysis comparing individual functional topography with a 30,000-subject normative atlas.

neuroimagingfmrihpcpnas-2026
§ 04

Thinking about

01Idea

Maslow had it backwards (sometimes).

The pyramid says you need food, shelter, money, love before you can do spiritual work. But a lot of the best art came from people whose base needs were not met — and arguably because they weren't.

02Idea

Why can't I search new places mid-route?

The real user task is "I'm going to X, but I also want to check out Y" — a near-universal trip pattern that today's maps still don't make easy.

03Idea

An AI-native instrument for music

Hardware: a foldable, Switch-sized core — a 12-key octave plus collapsible drum pads. Software: an AI music agent — hum a melody or describe what you want, and it arranges and lays it onto your tracks.

§ 05

Elsewhere

§ 06

Listening to

罗永浩的十字路口 For knowing interesting souls
Lenny's Podcast For how current products should work
§ 07

Say hi