A Personal Data Project
Your Saved Posts Are a Diary You Never Wrote
11,323 Instagram saves. Three years of private attention. An AI that reads every post, watches every video, and publishes an editorial magazine about the person who saved them.
11,323
Posts Analyzed
516
Gaza Events
760
Recipes Extracted
225
Person Profiles
21
Sideshifts
114
Training Exercises
5
Deep Dives
What Happens When AI Reads Your Archive
Instagram gives you a grid. No search. No analysis. No way to understand what you've been collecting. Your saves sit there — thousands of posts you tapped a bookmark icon on, then never saw again.
This project takes them back. Gemini 2.0 Flash watches the full video — motion, transitions, visual meaning that frames alone can't capture. Groq Whisper transcribes every audio track. Claude Sonnet reads all text sources and writes a searchable synthesis. Claude Opus reads 16 key frames per video and extracts every ingredient, every step, every word. Four AI models, each doing what it's best at, turning raw saves into something you'd actually use.
Recipes extracted from the food collection
A boolean flag from vision analysis found 97 recipes. AI agents reading all five text sources per post found 612. AI agents that watched the cooking videos found 760. A 684% improvement — not from better data, but from better seeing. A silent video of a chef slicing fennel became a complete recipe with 7 ingredients and 8 steps. The data was always there. You just needed an AI that could watch.
“Likes and comments are social behavior. Saves are private intent. This dataset is the closest thing to a real signal of curiosity, concern, taste, and identity.
Five Deep Dives
Each collection becomes its own editorial project — not a filtered dashboard, but a purpose-built narrative with its own voice, analysis, and design.
6,861 posts
Gaza
28 months of a war, week by week. 516 timeline events, 225 person profiles, 5,700 factual claims. A personal record of following a crisis through the accounts you trust.
1,415 posts
Counterculture
Anti-establishment politics and protest culture. 27% satirical content, a post-inauguration anger spike, and veganism as the fastest-growing theme.
851 posts
Food
8 years of recipes turned into a cookbook you’d actually cook from. 760 structured recipes extracted by AI agents that watched the cooking videos.
202 posts
AI
The fastest-growing collection — 6.5x monthly growth. Builder tutorials and surveillance critiques arriving simultaneously.
114 posts
Hundetrening
Positive reinforcement dog training from 95 accounts. Scandinavian mushroom hunting, consent-first philosophy, and 92% video — a learning journey made computational.
What the AI Sees
The Emotional Packaging Gap
53% of posts score positively on language sentiment — yet 57% are dominated by disgust, anger, or fear. Creators wrap difficult material in accessible, even optimistic, language.
Cross-Collection Sideshifts
Tucker Carlson is coded as fear in one collection and anger in another. 21 figures change emotional register depending on context. The archive contains the same people playing different roles.
The Rhythm of Attention
Saves peak at 4 PM in focused 8-minute bursts of 1.6 posts — punctuated by occasional 23-post deep dives that last hours. Private behavior, made legible.
The Filter Bubble Paradox
High source diversity (Shannon entropy 0.87) but extreme topical concentration — 86% converges on political content. Many voices, one concern.
Is This the Future of Software, Agencies, or a Mix of Both?
Until now, software and expertise have been two distinct things. You could buy software to process and automate complicated processes in a deterministic way — great for repeatable workflows. But to get expertise, analysis, and new ways of solving problems, you needed humans. The combination of non-deterministic creative expertise, data science, and technical processes has always been expensive and scarce.
This web presentation isn't the product. It's the result of a methodology: meticulous data science combined with editorial AI, all packaged into a Claude Code Skill that anyone with a Claude Pro account or above can use to create their own. The skill encodes every pipeline step, every gotcha, every type coercion rule. It's institutional memory, an agency, a data science team, technical helpers, and QA — in one tool.
This skill combines data science knowledge and strategies, libraries of scripts and methods, with a non-deterministic creative orchestrator that avoids processing data through a template. Instead, it takes any dataset, explores and exploits its uniqueness, uncovers hidden value, and builds real products from it.
One person and an AI engineering partner built five editorial deep dives, 760 structured recipes, 225 person profiles, and a psychological portrait — in a single day. The question isn't what we built. It's what you could build with the same methodology applied to your own data.
Explore
Overview
Collection metrics and trend summaries
Semantic Search
Find posts by meaning, not keywords
Post Galaxy
11,000 posts in semantic space
Topics
14 topic clusters and their evolution
Sentiment
Emotional geography of the archive
Network
690 accounts, 26 communities
Archive Profile
Behavioral fingerprint of attention
Browse
Explore every collection