Case Studies

Selected Work

Projects built with a focus on AI, developer tools, and real-world utility.

01 Open Source

ColdVPN

A one-click VPN that routes all of your Mac's traffic through a private server you own — so on public / coffee-shop wifi, the network only ever sees an encrypted tunnel, never where you're going. Zero manual setup: one command spins the whole thing up.

Under the hood, a virtual network interface captures traffic at the packet level, wraps it in a double-layered WireGuard path, and forwards it to your own exit server — every app protected, always on.

WireGuard GoGo Networking Swift
View on GitHub
coldvpn · tunnel
Mac
WireGuard
Your Server
Internet
encrypted · WireGuard tunnel
Timeline Q2 2026
Type Systems · Networking
Status Open source
02 Private

ColdSpot

ColdVPN with zero servers to rent: your phone becomes the relay and a free cloud VM becomes your private exit — one-click setup, nothing manual, and infrastructure you fully control end to end.

A deep networking build — Layer-3 packet capture, a userspace TCP/IP stack, a reverse tunnel for NAT traversal, SOCKS5-over-TLS with certificate pinning, and self-healing services. Kept private and shared with developers on request.

GoGo Networking TLS Terraform
Request access
coldspot · relay topology
Mac
iPhone
Oracle VM
Internet
encrypted · SOCKS5-over-TLS
Timeline Q3 2026
Type Systems · Networking
Status Private · request access
03 Active

ClassRec

A live lecture transcription SaaS that captures, transcribes, and organizes classroom audio in real time. Students focus on understanding instead of frantic note-taking. Built around OpenAI Whisper with a FastAPI backend and React frontend.

Supports real-time streaming via WebSockets, speaker diarization, and auto-formatted notes export. Designed with students as the primary user.

Python FastAPI OpenAI Whisper React WebSockets PostgreSQL
GitHub · Demo — coming soon
classrec.app
Live Transcript
REC
Professor: "Today we're covering neural networks..."
The architecture consists of layers...
Each neuron receives inputs weighted by...
Backpropagation adjusts these weights|
12:34
Timeline Q4 2024 — Present
Type SaaS · AI Tool
Status Active development
04 Shipped

MBGPT

An AI YouTube comment assistant that replies in the creator's own voice. Context-aware, human-toned, never generic. Demonstrates the evolution from baseline GPT to RAG-grounded to fine-tuned response quality.

Uses retrieval-augmented generation with video transcript context and fine-tuning on historical comment–reply pairs. Built as an end-to-end ML project from data collection to deployment demo.

Python OpenAI API RAG Fine-tuning Jupyter Vector DB
View on GitHub
mbgpt · comment assistant
This tutorial really helped me understand overfitting! 🔥
01 · Generic

Thank you for your kind comment!

02 · RAG

Glad the dropout explanation clicked for you!

03 · Fine-tuned ✦

haha yeah overfitting got me too for the longest time 😅 stay tuned next week

Timeline Q3 2024
Type ML · NLP
Status Shipped