Each project is written up the way a field guide describes a creature: where it was found, how it behaves, what meeting it taught me. Thirty seconds per entry is a fair reading. Twenty minutes works too.
specimen 01
This website alive
first seen
2026, after nine rounds of research that were partly wisdom and partly hiding
habitat
plain HTML, hand-written CSS, one small script. Under 100 KB, no trackers
behaviour
organizes itself by personal taxonomy; refuses templates; keeps its errors public
status
alive and expected to change
The most honest thing I can tell you about this portfolio is that making it nearly became a way of not making it. I generated a hundred concepts, audited them three times, tested them against a ten-year survival rule, and built a six-hundred-principle design genome, all before writing a single page. Some of that was rigor. Some of it was fear with a bibliography.
the longer reading — what the process actually was
The research produced five principles that survived every audit: organization reveals thinking; show the unfinished; give the viewer something; keep contradictions; make care visible in small details. Everything you're looking at is those five sentences, executed.
What surprised me: the audits kept killing my favorite ideas and sparing the plain ones. The concept I mourned most (an elaborate fictional archive) failed the simplest test: it impressed people without letting anyone meet me. The audit that did the killing is kept, verbatim, in 004.
What I'd do differently: fewer rounds, earlier drafts. The tenth version of a plan is worth less than the second version of a thing.
specimen 02
The rule-leakage experiment resting
first seen
2026, a Python simulation built out of curiosity about AI systems
habitat
a simulation script, two rule files (one public, one private), and a results pickle
behaviour
tests how much of a hidden rule set can be inferred from a system's visible behaviour
status
resting: results collected, conclusions not yet written up
A small experiment about a large question: if a system follows rules you can't see, how much do its answers give away? I built a toy version — public rules, private rules, and a probe that measures the leak. ask me about this
the longer reading
What I found, what surprised me, and what remains unresolved belongs here. The honest current state: the results file exists, the write-up doesn't. That gap says something about how I work: collection comes easily, conclusion takes discipline.
specimen 03
Automations that outlived their novelty in use
first seen
2025–2026, workflow experiments: n8n pipelines, research agents, call-automation studies
habitat
the unglamorous middle of my computer
behaviour
quietly does the same boring thing correctly, every day
status
in use, which is rarer than impressive
Not portfolio pieces in the usual sense. But a tool that a person actually uses every week says more than a demo that was used once, for a screenshot. The one I'd show you first is the research briefing pipeline.
specimen 04
The trading signal that wasn't finished
first seen
2026, a claim I kept reading on Indian financial Twitter and decided to take literally
habitat
Python, pandas, statsmodels; eight years of NSE market and institutional-flow data; a five-stage pipeline
behaviour
predicts nothing, and keeps a full record of how it found that out
status
finished, and less flattering than its first draft
I tested whether the disagreement between foreign and domestic institutional money predicts which way a sector index is mispriced. It doesn't: zero of sixteen sector and horizon combinations survive correction for multiple comparisons. My first version said otherwise. It was wrong in three separate places, and I found all three by auditing my own work, which is the part I would rather be judged on. The longer reading is the write-up itself.
the three ways it was wrong
A holdout set that silently merged back into training during a routine data refresh, which made every out-of-sample claim in-sample. A headline statistic in the write-up that no longer matched the CSV sitting next to it, because the data was regenerated and the document wasn't. And the finding I liked most, a tidy structural story about futures positioning, built on a hundred days of data and a comparison between a level and a flow. Redone properly on the full 1,750-day overlap, it turned out not to be backwards, just irrelevant.
All three are documented in the repository rather than quietly fixed, on the theory that the corrections are the part worth reading: capm-flow-predictor.
the unfinished
Things I want to build and haven't
Listed in public, because a wish written down behaves differently from a wish kept in my head.
The rule-leakage write-up, done properly, with the humility the data deserves.
A tool that turns my reading into a searchable personal library.
Something physical. I have spent too long inside the screen.
placeholder page-note: specimens 02–03 need your real details where marked ＊. Specimens 01 and 04 are fully real and need no editing: one is this site's own history, the other is a finished research project with its write-up attached.