← All projects
Experiments
Side Builds
2025 – 2026
Side Builds

Built because curious.

01

MAX

What if AI agents could argue with each other?

Most AI tools are single prompt, single response. You ask, it answers, you move on. MAX is different.


Four specialist agents with distinct expertise enter a structured 3-round debate about financial risk scenarios. One proposes. Three challenge in parallel. The first rebuts. What comes out is sharper than any single agent could produce.


The finance domain is just the sandbox. The real point is the orchestration pattern: peer-to-peer agent debate as a protocol. Agents don't just answer prompts. They challenge each other's reasoning, find blind spots, and build on disagreements.


Each agent has its own "soul" (system prompt), memory file, and inbox. They communicate through a file-based message routing system. The orchestrator manages the debate lifecycle, tracks risk scores across 6 domains, and archives results for future reference.

Protocol
3-round structured debate. Round 1: Max proposes analysis. Round 2: CPA, Cross-Border Tax, and Tax Attorney critique in parallel. Round 3: Max rebuts with synthesis.
Agents
4 specialists with distinct system prompts: Max (generalist), CPA (compliance), Cross-Border Tax (international), Tax Attorney (legal risk).
State
File-based inbox pattern. Each agent reads from and writes to its own inbox. Persistent risk scores across 6 domains (FBAR, FATCA, Crypto, GILTI, Form 5471, PFIC).
Architecture
Spawns Claude CLI as child processes. Parallel execution with 180s timeout. Graceful fallback. Process lifecycle management with SIGINT/SIGTERM cleanup.
Debate Protocol
MAX Proposes analysis across 6 domains ROUND 1 PROPOSE CPA compliance CROSS-BORDER TAX TAX ATTORNEY ROUND 2 PARALLEL CRITIQUE MAX Rebuts with synthesis Updates risk scores ROUND 3 REBUTTAL + SYNTHESIS
Bun TypeScript Claude CLI Multi-agent File-based state Parallel spawning Source →
02

XHS Outreach Bridge

Automating on a platform that fights back.

Xiaohongshu (Little Red Book) is one of the most aggressive anti-scraping platforms in the world. Real-time behavioral analysis. Device fingerprinting. Typing pattern detection. Most automation gets caught in minutes.


This tool bridges two platforms: Qian Gua (a third-party analytics platform) for influencer discovery, and Xiaohongshu for outreach. Find targets on one, contact them on the other, while staying under the radar.


The challenge isn't writing the code. It's thinking like a human. How fast does a real person type? How long do they pause between actions? Do they scroll before clicking, or click immediately? Every behavior pattern is a signal, and XHS is watching all of them.

Human Simulation
Keystroke variation, not just delays. Typing speed varies per character. Pauses between fields. Scroll behavior before actions. The goal is to be indistinguishable from a bored person browsing.
Anti-Detection
Single-tab discipline. Multiple tabs = instant flag. Stop-on-red-flag protocol: detect anomaly, wait, retry with different timing, notify operator if blocked.
Data Bridge
mitmproxy for share link conversion. Bulk convert 300+ discovery URLs to xhslink.com share links. iPhone Mirroring integration for mobile-native flows that bypass desktop detection.
Auth
CDP-based browser automation. Cookie extraction is broken on macOS for XHS. CDP (Chrome DevTools Protocol) connects directly to an authenticated Chrome profile instead.
Platform Bridge
QIAN GUA Analytics platform Influencer discovery Performance data BRIDGE LAYER URL → Share Link Conversion Human Speed Simulation Behavioral Pattern Matching Red Flag Detection mitmproxy CDP iPhone Mirroring XIAOHONGSHU Comment outreach Profile engagement Anti-scraping hostile ⚠ Single tab only. Stop on red flags.
CDP mitmproxy iPhone Mirroring Xiaohongshu Qian Gua Anti-bot evasion Source →
03

Rolling Damage Report

Three pivots. One viral launch. Zero ad spend.

Started as a body journaling app. Track your injuries, log your recovery. Nobody cared.


Pivoted to BJJ-specific body journaling. Closer. People in the Brazilian Jiu-Jitsu community got the problem. But they didn't want to journal. They wanted to roll.


The community kept asking the same question: "Where can I find open mats near me?" No good answer existed. So the app pivoted again. Became an open mat finder.


Posted on Reddit. It went viral. 60+ gyms submitted their open mat info within weeks. All organic, all free, all because the product finally matched what people actually wanted.

V1
Body Journal
General injury tracking. No traction. Too generic, no community.
↓ listened to signal
V2
BJJ Body Journal
Niche down to Jiu-Jitsu. Interest grew. But journaling wasn't the core need.
↓ listened harder
V3
Open Mat Finder
Built what the community was actually asking for. Reddit launch went viral. 60+ submissions.
Launch
Reddit viral post. Shared in BJJ communities. Immediate organic traction. No paid promotion, no influencer outreach, just the right product for the right audience at the right time.
Traction
60+ organic submissions from gyms across the US. Community-driven data. People submitted their own gym's open mat schedules without being asked twice.
Product
1,987 academies mapped. Real-time interactive map with clustering. Searchable by state and city. Community submission form with admin approval flow.
Live
rollingdamagereport.com — shipped and running. Monthly auto-redeploy. Active community submissions.
0
Pivots
0
Organic Submissions
0
Academies Mapped
Next.js 16 Convex MapLibre GL Supercluster Vercel Reddit GTM Source → Live Site →
Built by

Jenny
Chen

GTM operator turned builder. These are the side experiments. The case studies live on the main portfolio.

Side Builds
Experiments
2025 – 2026