Ethan Lung Portfolio

01

AccessWayfinder

Accessibility · iOS

An iOS app that rates local businesses on accessibility — entrance, interior, parking — so people with mobility needs know before they go. Started in Westfield, NJ. I led the 12-person team that shipped it to the App Store.

Problem
Older buildings in town don’t always meet ADA standards, and there was no way to know before visiting.
Data
Visited and called 500+ local businesses. Each is rated 1–5 on entrance, interior, and parking, with notes on restrooms and seating. Users add their own ratings.
Build
React Native and Expo. Supabase for storage and the photo pipeline. City browsing and category filters.
Ship
App Store release through EAS, then more towns: Cranford, Summit, Princeton, Jersey City.
Press
Covered by PBS (NJ Spotlight News) and NJ.com.

stackReact Native · Expo · Supabase

users: 3,000+
App Store rating: 5.0

App Store PBS feature

Downtown Westfield, NJ

500+ businesses rated · hover one

Show Click a pin to keep its card

02

SocialPhysics

Audio · Networks

Listens to a group conversation, identifies who is speaking, and scores each person’s influence in real time with an “Attention Economy” algorithm. Speech events are classified in under a second. The result is drawn as a force-directed graph of the speakers.

PipelineResemblyzer
Google Speech-to-Text
Streamlit
Pyvis

Source

Hover a speaker

Stylized version of the live Pyvis graph

Click a speaker to make them talk. Influence moves toward whoever holds the room.

03

TennisTrades

Tennis · ML · Markets

A trading engine for live ATP matches. An XGBoost model estimates the win probability as each point is played; when it disagrees with the Kalshi price, the engine places limit orders in under a second.

Each pointlive ATP matches
XGBoost model
Kalshi price
limit orders in under a second

matches modeled: 5,000+
prediction accuracy: >75%

Source

Win probability: 81.0%
Kalshi price: 0.76
Model edge: +5.0%
Surface: Hard
Shot: 10 / 10

Model vs. market, shot by shot Hover to scrub

XGBoost win probability Kalshi price Edge Illustrative point, not real data

Point Surface

Amazon · New York, NY · Summer 2026

Software Development Engineer Intern

Built tooling that unifies fragmented cloud-network data into one topology and answers “would this packet get there?” across a large AWS environment

01 Collect
A read-only collector pulls network data from hundreds of accounts every 30 minutes.
02 Model
One graph snapshot per scope. Every reader is served from it, with zero live calls on the read path.
03 Trace
Hop-by-hop longest-prefix match over the collected routes: delivered, no route, or blackhole, with the matched route at each hop.
04 Serve
Two APIs, get-topology and trace-route, behind a UI and an MCP server I built so the team’s AI agents can call them as tools.

Shipped with alarms, integration tests, and a runbook.

300+
accounts: 8
regions: 600+
route tables: 0
write permissions

Before20+ minutes

After<2 seconds

a connectivity investigation that meant reading docs, DNS logs, and route tables by hand across accounts, now one query

Traces run left to right · hover a source to trace it

Visualization is synthetic and does not represent Amazon infrastructure.

From To

Jirav · Series B · Jan – May 2026

Software Engineer, AI Systems (contract)

An AI analyst inside Jirav’s FP&A platform: ask in plain English, get an answer from live financial data

7
tool calls per query: 9
tools: 4
workflows

Validated across 4 workflows with finance professionals at J.P. Morgan, BCG and EY.

“What’s our runway if we hire three engineers in Q3?”

resolve_context90 ms
fetch_actuals210 ms
fetch_forecast220 ms
fetch_headcount180 ms
query_financials240 ms
calc_burn60 ms
calc_runway70 ms

Illustrative trace. Three fetches became one batched read and the calculators were composed; 9 tools in total.

Architecture

Nue.io · Series A · Sep 2025 – Jan 2026

LLM & Software Engineer (contract)

Pricing logic written in plain English, shipped as validated Salesforce CPQ plugins

90%
faster plugin development: 40%
fewer logic regressions: 25+
production plugins

Plain English

Apply a % discount when ARR exceeds $k.

RuleIF arr > 100000 THEN discount = 0.20
Grammarparsed, validated against the plugin grammar
JavaScriptreturn arr > 100_000 ? 0.20 : 0;
Deployplugin → production, gated by CI validation

Off court.

I’m Ethan, an EECS student at Berkeley interested in software, systems, and building things I’d actually want to use.

Outside of that: tennis, music, golf, poker, and trying different foods.

Degree: B.S. EECS, UC Berkeley · GPA 3.54 · May 2028
Recognition: Special Congressional Recognition (Rep. Kean) · Congressional App Challenge · Featured by PBS & NJ.com

Looking for Summer 2027 software engineering internships: systems, infrastructure, or applied ML.

ethanlung@berkeley.edu

Ethan Lung.

Off court.

Ethan Lung