This Summer, I Hired College Kids to Clone My Brain
Philly AI Lab brought together five college students around one strange assignment. They built custom systems around my news, writing, analytics, and relationships. The more accurately those systems began to reflect me, the harder it became to ignore the question underneath them: who owns the memory, voice, and judgment they learn to carry?
This summer I launched Philly AI Lab and hired five local college students to build what I jokingly called my second brain. By August, the joke had become literal. They had built custom software around core foci of my work: understanding what is happening in AI, learning what is working across my content and what I might want to say next, and searching the 5,000+ people in my network when somebody needs a job, a candidate, an expert, or an introduction.
I am using the apps now, which matters because this was never meant to be a summer exercise that ended with a presentation. Philly AI Lab was set up as a production-grade summer residency, and an investment in myself and my business. We began with decisions I make repeatedly, looked at the information I need to make them well, and built software around the way I already work. That specificity helped ensure the apps remained grounded in my day-to-day life.
This said, the apps are still messier than the phrase “second brain” suggests. The system spans three codebases, several models and APIs, Airtable, Ghost, Pinecone, and different stores for news, writing, analytics, and relationships. It’s not a synthetic Christie in one monolithic database, but the students built pieces of infrastructure that could extend my mind, retrieving context I usually struggle to handle on the fly.
The more useful the system became, the less interested I was in whether we had actually cloned my brain, and more interested in how to keep the system specific enough to remain an extension of me while the models underneath it keep changing.
What the residents built
Mazen, a junior at Drexel, built the news engine. It pulls from roughly fifteen AI-news feeds, embeds and ranks the articles, clusters related stories, generates short editorial captions, and checks the claims in those captions against the original reporting. The result is a smaller, ranked view of the developments most likely to matter to the work I am already doing.
Kush and Danny, both incoming college freshmen, built the dashboard that sits on top of that feed. The Performance tab pulls my LinkedIn, Ghost, Substack, and Medium analytics from Airtable so I can see growth, engagement, and which pieces are actually landing. The Ideas tab reads Mazen’s ranked news alongside my published writing and performance history, then suggests possible topics, angles, hooks, and editorial moves.
The inspiration and design here directly tie into my weekly workflow. Given what is happening now, what I’ve already said, and what my audience has responded to before, what might be worth saying next? The assistant can help turn a dictated idea into a draft, but it is also designed to keep my writing separate from outside news, cite what sources a concept uses, and effectively ground all of its outputs.
These nuances matter if the system is going to suggest what I should post about. First it needs to know which words are mine, which facts came from outside reporting, and what is just noise.
Jie, a recent Temple graduate, and incoming Villanova master's student, worked on the relationship side of my brain – more specifically creating an AI-native CRM tool that I can use on the fly. From my phone, I can record a voice note, upload a business card or résumé, scan a LinkedIn QR code, or add a contact manually. The multimodal app extracts the contact’s information, prompts me to review, enriches the missing fields (upon request), and saves the connection into Airtable.
Beyond adding contacts on the fly, there's a second feature of the app that is critical to my day-to-day as a superconnector of people – ai-native search. For instance I can ask the search something like “I met a computer science student looking for a job; who could help them?” Unlike traditional search it does not simply return other students or people with computer science in their profiles. Through a multi-step process identifying keywords, fuzzy matching, and then agentic search (in order to maximize relevance and minimize latency) it able to identify recruiters, hiring managers, connectors, and other people in my network who could actually help, then explains why each person may be relevant.
A lot of people come to me when they are looking for work, trying to hire, raising money, searching for an expert, or simply trying to meet the right person. I love connecting people, but I cannot reliably carry 5,000+ relationships in working memory. I wanted “Who do I know who could help with this?” to become something I could ask the system instead of something I try to remember off the cuff.
Miya, our strategist, worked across the system rather than inside one project. She ran the content calendar and LinkedIn page, shaped the story around the builds, and made sure the technical work stayed connected to an actual operating rhythm instead of becoming four uncoordinated projects.
Together, the projects answer three questions I ask constantly. What is happening? What might I want to say about it? Who do I know that could help? To me, this makes for a useful second brain that delivers context at the moment a decision needs it.
Why custom matters
Most second-brain products begin with a container. You move your notes, files, meetings, and habits inside the product, then learn to work the way the product expects. We started in the opposite direction. The news engine owns news. Airtable owns structured analytics and relationship records. My social media accounts remained the source of truth for my published writing. The dashboard reads across those systems without treating them as interchangeable.
Those boundaries are technical, but they are also how I began defining authenticity in the system. The dashboard does not get to treat every sentence it retrieves as mine. The CRM searches actual relationship history rather than inventing a connection. The news engine checks its captions against the source before passing them downstream. Each application has to know what it knows, where that knowledge came from, and when it should stop.
I wanted a custom system because I did not want a generic AI guessing at who I am from a few instructions and a folder of documents. These applications are grounded in the actual record of my work, and I can inspect, change, or remove the context behind them. That gives me more control, but it does not eliminate the role of the model sitting between my ideas and the finished output.
That distinction becomes more critical as frontier models improve. When the output is clumsy, authorship is easy to see. When a model can produce a polished argument, mimic a voice, and move between research, writing, and code, the boundary between assistance and substitution becomes much harder to feel from the surface.
I do not believe authenticity requires me to type every sentence or write every line of code. I do believe it requires that the ideas, direction, judgment, and responsibility remain my own.
When the model starts shaping the voice
I had been thinking about that already when Anthropic explained how it plans to watermark text from future Claude models. The watermark does not add a hidden character. It uses low-stakes word choices to create a statistical pattern that can later suggest Claude was involved in the text.
Anthropic says the method has no practical effect on quality and does not change ownership. I understand the regulatory problem they are trying to solve, but it still gives me an ominous feeling. We need better authorial standards for AI-assisted work, a meaningful definition of “AI slop,” and better ways to deal with genuinely malicious uses of AI.
But a company I do not control, altering the layer between my thoughts and their expression to serve a forced government objective, seems to miss the point… and compromise the output deliberately for the purpose of tracking it. Providers may argue the difference in quality is negligible, but my question is why a European regulatory objective should carry any weight aan author’s creative freedom and intended expression?
The issue is not that Claude participated. I use AI throughout my work. The issue is that participation becomes the signal even though it tells us very little about authorship.
We have spent a long time confusing the craft of production with the value of the idea. I have an art degree. I understand the instinct. Difficulty is not the same thing as value, and using a more powerful tool does not automatically make the thought less yours.
If I know nothing about Python, ask a model to write an article about Python, barely understand the result, and put my name on it, I am not meaningfully the author. If I develop the ideas, understand the material, direct where it goes, exercise judgment over what stays, and accept responsibility for the result, authorship should not depend on whether I personally assembled every sentence.
This conversation will be the focus of the next issue of Superabundance: the minimum viable author. I want to get more precise about the point at which AI remains a tool in the creative process and the point at which the human has stopped doing enough of the intellectual work to claim authorship.
For this project, the same question appears in architectural form. The new frontier models are becoming good enough to serve as the interface through which I research, write, connect, and build. If the memory and voice that make the system mine live inside the provider too, then the provider is doing more than supplying intelligence. It is mediating the version of me that comes back. This is increasingly my greatest concern.
Comparing notes with Luis
On Friday I talked with Luis from The Hooman Loop, a fellow Philly tech friend who has been building a business brain from the other direction. He runs Hermes as an orchestrator on a local Mac mini and talks to it through Slack and Telegram. It has its own carefully limited accounts and permissions, and for software work it can move between Linear, Cursor, GitHub, and pull requests before reporting back to him.
Much of the durable knowledge sits in ordinary Markdown files: the business model, personas, positioning, testimonials, lead magnets, preferences, workflows, and the rules the agent should learn over time. His most consequential architectural choice is to treat the durable context as a separate layer from the model. The model can change without taking the memory of the business with it.
My students built applications around the way I work. Luis built a layer designed to keep the context separate from any one model. His setup made the chain much clearer to me. Custom software can reflect me more accurately; owned and portable context gives me a better chance of keeping that reflection mine when the frontier model changes underneath it.
That is the direction I want to move. I do not need my stuff to run locally on a mac mini, but I do need more of the durable memory, voice context, and orchestration under my control so I can change models without rebuilding the accumulated version of me that makes the system useful.
Over time, that may mean more self-hosted components and a more deliberately fine-tuned brain and content system whose behavior I can inspect and evaluate against my own standards. Ownership will give me greater ability to move, correct, and govern what the system thinks it knows about me.
How to build your own
Most people can start with far less than three custom applications and an agentic software-development lifecycle. The useful starting point depends on how specific the problem is.
1. Start with a bounded project
The fastest route is to put a defined set of context into a tool you already use. ChatGPT Projects and Claude Projects can keep files, instructions, and an ongoing body of work together. NotebookLM is useful when source grounding is the point. These are low-friction ways to learn which kinds of context are actually valuable before you build anything.
2. Use a structured knowledge workspace
I looked more closely at Mem and Tana before putting them in this issue because I have not used either long enough to recommend either one. Mem is the more convenience-oriented option: capture notes, meetings, voice, and research, then let an AI agent use that context later. It can export notes and collections to Markdown, although the product still lives in Mem’s cloud. Tana Outliner is a more configurable knowledge graph built around nodes, Supertags, saved views, voice capture, and custom automations. Both are credible middle paths. I have not used either long enough to tell you which one earns a place in your life.
3. Build around a specific decision
Custom starts to make sense when the questions are specific, recurring, and valuable enough that generic memory misses the point. The CRM had to tell me who might help a particular person, based on evidence and relationship strength, rather than merely remember everyone I had met. The content system had to bring ranked news, my own performance history, and my writing corpus into one editorial workflow rather than stop at summarization.
Build the smallest system that answers the question you actually have. Add an agentic layer when maintaining that system becomes the next bottleneck.
I started the summer asking how much of my work I could hand to AI. The better question turned out to be what I had to make explicit about my work before a machine could help.
The students had to identify what I pay attention to, how I judge relevance, what makes a connection useful, what counts as evidence, and where my voice ends and outside information begins. Luis made me think about the layer above that: how the context survives changes in models, tools, and people.
A second brain can remember. A personal operating system can apply that memory. A system that is meant to last also has to survive the people, models, and vendors that built its first version.
The next phase is teaching mine how to change without losing what made it mine.
What you should read
Five reads on the harder half of this, which is not building the AI but getting real value out of it once it exists.
1. Understanding the AI Economy
Zanna Iscenko and Scott Strand, Google, July 23, 2026 · AI economy, adoption, real-world usage
The biggest real-world look yet at how AI is actually used. Google analyzed fifteen million anonymized human-AI conversations and found adoption is wide but shallow: AI touches most jobs, but only a slice of the tasks inside them, and it rarely does the whole job.
- AI showed up in about 21 percent of work tasks, across roles that cover most of the economy.
- Fewer than one in ten of those uses fully automated the task; most were collaboration, not replacement.
2. The Anthropic Economic Index
Anthropic, June 26, 2026 · AI at work, augmentation, worker outlook
The companion dataset from the other side, built on how people actually use Claude at work. The through-line is that most AI use speeds people up and widens what they can take on rather than replacing them, and the workers leaning in hardest are the most optimistic about their own future.
- 86 percent say AI speeds their work and 82 percent say it widens what they can do; 57 percent think it raises their skills’ market value.
- More than a third expect AI to handle most of their tasks within a year, so this augmentation-heavy picture is a snapshot, not a settled state.
3. From adoption to impact: Three horizons of AI transformation
McKinsey, July 2026 · transformation, workflow redesign, organizational readiness
McKinsey splits AI maturity into three stages: enablement, automation, and reinvention. The finding that matters is that handing people tools is not enough, because individual productivity rarely becomes company value unless workflows, roles, and operating models change too.
- Only about one in ten leaders said their organization had reached the reinvention stage.
- Organizational readiness explained more of the value gap than individual readiness; the tools alone do not get you there.
4. Research: How AI Agents Broaden the Scope of Knowledge Work
Jeremy Yang, Kate Zyskowski, Noah Yonack, and Jerry Ma, Harvard Business Review, July 2026 · agents, knowledge work, judgment
A clean way to see what changes when agents enter the picture. Old tools hand you information to act on; an agent hands you finished work, which pushes the human job toward deciding what is worth doing and whether the output is any good.
- Agents do not just speed up knowledge work, they change its nature, from doing the task to reviewing and directing it.
- The scarce skill becomes judgment: framing the problem, checking the result, and deciding what deserves to ship.
Who you should follow
A few people worth following on building and living with an AI brain, from the engineers who make the models to the people who have thought hardest about notes and memory.
Luis Cielak · founder, The Hooman Loop
Luis is the person I compared notes with in this issue after he showed me the local, modular second brain he built to support his business. At The Hooman Loop, he works across product strategy, UX, workflows, and AI-native systems, helping founders and teams make hard decisions when the work is ambiguous or expensive to get wrong. He also leads Cursor Philly, a 750+ member community for AI-assisted development.
Chris Alexiuk · AI research engineer, NVIDIA
One of the clearest explainers of how these models actually work, from retrieval to fine-tuning to agents, which is the machinery underneath any AI brain you try to build.
Allen Downey · data scientist; author of Think Python and Think Bayes
Writes about data and probability the way this issue thinks about AI: start from how you actually reason, not from the tool. His blog, Probably Overthinking It, is a quiet antidote to hype.
Tiago Forte · founder, Forte Labs; author of Building a Second Brain
He wrote the book on the second brain, and has spent the past year working out what changes when the brain is an AI. The obvious first follow if this issue got you started.
Anne-Laure Le Cunff · founder, Ness Labs; neuroscientist and author of Tiny Experiments
For the other half of the question, how notes and memory actually work in a human head, from someone who studies the neuroscience and runs one of the best communities on the topic.
What’s moving in Philly
A handful of rooms worth being in around Philly over the next couple of weeks, from data-and-AI happy hours to hands-on coding sessions.
Philly Data & AI: August Happy Hour
Aug 25 · Con Murphy’s Irish Pub, Philadelphia
Drinks and real conversation with the Philly data and AI community, no speakers and no sales pitches.
Startup Bucks Networking Happy Hour
Aug 27 · Crossing Vineyards and Winery, Newtown
A low-key monthly gathering for founders and entrepreneurs out in Bucks County, drinks and conversation, no agenda.
DataPhilly Tech Talks
Aug 27 · Slalom Consulting, Philadelphia
Talks and networking from the Data Science Philadelphia crew, for anyone who wants substance with the evening rather than just a mixer.
Cups of Code: Connecting the Community
Aug 30 · Panera, Philadelphia
An easy afternoon of developers coding side by side over coffee, good if you would rather build than network.
Claude Code Philly
Sep 5 · Turkish Brew, Philadelphia
A weekly session where developers pair up and build with Claude, hands-on and low-key, and a gentle way into September.
(powered by Kynra)
Copyright 2026 - Christie Mealo