Puzzle Exchange
Human first tech community with 6k subscribers.
A panel series in San Francisco. Pick a room.
Twenty four channels in. One screen out.
One person, not one inbox per app. Twenty four channels resolve to a single thread and a single history.
Ask in plain words. It reads your files, messages and accounts directly, and cites what it opened. The store is yours, not a vendor's index.
Answers come back already written from the real account history, sourced and ready to edit.
It can draft. It cannot speak for you. Nothing reaches a person without your tap.
Each one started with a process that was too manual, too slow, or too dumb to keep doing by hand. Click through for the full story.
A personal operating system that consolidates 24 channels into one database and one approval gate. Every proven flow becomes a named verb: a deterministic call that costs zero tokens, which is what lets the loop run every fifteen minutes instead of once when you remember to ask. Nothing sends without a human grant.
Chrome extension that turns Luma event guest lists into LinkedIn connections automatically. Reverse-engineered LinkedIn's internal Voyager API. Launched at SXSW and Shoptalk. Real users, real complaints, constant platform adaptation.
Seeding Ninja maps any D2C category in 48 hours. Searches Reddit, TikTok, Instagram, and web to surface competitors, pulls their full influencer collaboration history through Modash API, and identifies which creators actually drive growth. First run: fine jewelry. Same system works on any vertical.
AI powered referral discovery for an acupuncture practice. Multi-channel outreach across Instagram, Gmail, and events. The most complete build: live campaign, real results, full operations stack.
A non-destructive edit engine for long-form video: one source, one content-addressed segment cache, and any number of variants off it. Changing a single cut used to re-encode the whole program. Now only the touched segments rebuild, which is what makes iterating on a 75-minute episode practical instead of overnight.
A terminal pipeline that takes a raw panel or podcast recording to a broadcast-clean long-form master: denoise, cut the dead air, then two-pass loudness mastering to spec. Each stage dispatches to a swappable tool, so a better denoiser drops in without rewriting the chain around it.
Not just building with AI. Building frameworks for how to build with AI. Each is a Claude Code skill used across all projects above.
Original framework built from neuroscience. Every interface is a prediction error decision. 5 cognitive layers, 8 failure modes, 3 output formats.
Reverse-engineered from Claude's system prompts. 14 sections, 12 common mistakes, 4 prompt types. Used to build every agent in this portfolio.
Catches AI tells by structural pattern, not just phrases. Integrates brand voice matching for consistent output.
One slash command that survives /clear. Auto-detects SAVE vs LOAD, per-project history, lives outside the repo.
Patterns for keeping multi-agent runs under 30% main-thread context spend. Hard caps on handoffs, marker-line protocol, eight named anti-patterns.
Each catches a different
way the model drifts:
perception, prompt,
output, memory,
attention.
A panel series in San Francisco. Pick a room.
Projects built because the problem was interesting. Less polished, more fun.
4 specialist agents debate financial scenarios in 3 structured rounds. Built to explore peer-to-peer orchestration patterns where agents challenge each other's reasoning, not just answer prompts.
Automating outreach on China's most hostile anti-scraping platform. Human-speed typing simulation, behavioral pattern matching, and cross-platform data bridging between analytics and engagement.
Body journal → BJJ journal → open mat finder. Three pivots based on community signal. Reddit viral launch. 60+ organic submissions from across the US with zero ad spend.
GTM operator turned builder. I find processes that are too manual, too slow, or too dumb, and I build the AI systems that replace them. Targeting GTM + Partnerships roles at Series A–B AI companies.
