I analyze my own personal data as a habit, not because anyone asked, so when Instagram made a full data export available I built a tool to actually look at what was in mine. It's a single self-contained HTML file that reads the export ZIP entirely in the browser: nothing gets uploaded anywhere, and the tool is built to ignore message and comment content outright, extracting only timestamps and anonymized contact IDs.

From that stripped-down data it computes real sociability metrics: who initiates conversations versus who responds, how long each side takes to reply, how concentrated my messaging is across my top three contacts, and what share of my activity happens overnight. It's the same kind of passive behavioral signal I work with professionally on Beiwe, just pointed at my own Instagram history instead of a research participant's.

What's next: Next is extending the same sociability metrics (initiation ratio, response latency, nocturnal activity) to other personal data exports beyond Instagram.

Part of running operations for the Beiwe Research Platform at Harvard's Onnela Lab is keeping it funded, which used to mean checking Grants.gov, the NIH Guide, and half a dozen foundation sites by hand, hoping I didn't miss a deadline three weeks out. I built a Python scraper that pulls open solicitations from public grant sources and has Claude score each one against the platform's actual research profile.

The harder problem wasn't the scraping. It was realizing that grant solicitations are written in the language of the disease or clinical question a funder wants solved, never in the language of the platform that ends up solving it, and rebuilding the search around that instead of around better scoring.

"This R24 infrastructure mechanism directly aligns with BSC's mission to provide shared digital phenotyping research resources, though the NIGMS/ODSS focus is a softer fit compared to NIMH or NIDA."— Claude, scoring grant relevance

What's next: Next is rewriting the search vocabulary around disease and treatment language instead of platform language, then wiring results into a Notion database so a new run only surfaces what's actually new.

In 2018 I started tracking my own Apple Watch HRV data, just personal curiosity. When I exported the same historical stretch twice, once in 2020 and again in 2021, the numbers didn't match. Nothing about my actual heart rate history had changed, but Apple's algorithm had silently reprocessed it in between.

I shared what I found with JP Onnela, a Harvard biostatistician and the developer of Beiwe, the research platform I now run operations for. The finding made The Verge, and looking back, it's the moment I realized I wanted a career building the tools that collect health data properly, not just analyzing what other systems handed me.

"These algorithms are what we would call black boxes — they're not transparent. So it's impossible to know what's in them."— JP Onnela, The Verge

What's next: Less than a year after this made headlines, I became Head of Platform for Beiwe, the same research platform JP built. These days I get paid to do the thing I was already doing for free.

Beiwe, the research platform I run at work, used to be able to collect call and text metadata directly from a participant's phone, until Apple changed what third-party iOS apps are allowed to see. I got curious whether there was still a way to get at equivalent data some other way, starting purely as a personal project on my own Mac, with every identifier hashed from the very first line of code.

Passive call and text metadata turns out to be a genuinely rich proxy for sociability, and building this taught me more than I expected about two things: verifying an assumption before trusting it, and designing privacy protection around what an attacker could actually do, not just what looks secure on the surface.

What's next: Exploring whether a packaged, non-technical-participant version of this pipeline could ever become an official Beiwe workaround. Still just an idea at this point, not yet proposed anywhere.

I work at the intersection of academic research, software development, platform operations, technical support, and client account management — which means constant context-switching across very different kinds of work. Managing that manually, keeping it all in my head and improvising responses to recurring situations, was unsustainable. So I built what I think of as a professional operating system: a defined set of inputs, transformation rules, storage locations, and outputs that runs recurring work without me reinventing it every time.

The piece walks through six foundational principles — systems over silos, low friction, documentation, knowledge bases, automation, and AI as a transformation layer — and then applies them end-to-end: how Notion functions as thinking infrastructure, how Todoist stays a pure execution layer instead of a dumping ground, how meetings get prepped and closed out before the transcript is ever touched again, and how the inbox becomes an input stream instead of a to-do list.

"You shouldn't be the integration layer between your own tools."— Systems Not Silos

What's next: This is the first in a series on building a professional productivity system — upcoming pieces go deeper on knowledge base design, AI workflow architecture, and a CRM system that actually supports how you work.

Job searches are overwhelming — juggling resumes, tailoring applications, figuring out which roles actually align with a long-term goal. I wanted a tool that could stay organized, give strategic feedback, and support the whole process. So I treated it like launching a real product: built, tested, refined, and iterated on prompts to fix hallucinations — with my sister, who was job-hunting herself, giving real-time feedback on which features actually mattered.

My own background spans clinical research, operations, analytics, product management, and two startup stints — I needed something that could see the pattern across very different opportunities and gradually build a picture of what an ideal role actually looks like, not just score one job at a time.

"At any point, I've been able to ask JARVIS to show me an updated job ranking table and it's good about giving you a full list of all the jobs you've fed it, sorted by fit score and the details about those jobs."— JARVIS user

"One of the best things about the tool is that it keeps me focused on looking for jobs that align with the long-term plan JARVIS suggested. This has helped me keep my search targeted toward my own growth goals, rather than falling into that desperate mode where we apply to everything."— JARVIS user

What's next: Live job feed integration for proactive matching, ongoing skill-gap analysis, a built-in application tracker with interview prep, real-time ATS scoring, and LinkedIn profile alignment suggestions — building on the same foundation as the search itself evolves.

Fantasy football is part skill, part luck, and a whole lot of research — juggling analyst sites, injury updates, and endless flex debates. I built a persistent, stateful GPT that acts as a personal GM: it remembers my roster across the season, builds an ideal lineup each week from matchup-based projections, produces sourced player reports (ESPN, FantasyPros, Yahoo, RotoWire, Draft Sharks), evaluates waiver and trade moves with tier-based value, and proactively flags injury news on my own roster with replacement suggestions.

Structuring the role, defining outputs, and wiring in stateful memory so it evolves across a 12-man PPR league season was the real challenge — and it's already saved hours of weekly research.