Archive
A portfolio that only shows finished work is a partial account. These are real directories on a real disk. Where the reason something stopped is not recorded anywhere, the entry says that rather than inventing a story — an undocumented abandonment is itself an accurate thing to display.
The first attempt at proctored online exams. Its entire README is the two words "Hello, Git!" — the repository was initialised and then outgrown before it was ever described. A zip of it sits beside the folder, which is what saving something you have stopped believing in looks like. The idea survived; it became GreenGuardian, which solved it by moving the vision into the browser. This directory is the version that had not worked that out yet.
Same two words in the README — "Hello, Git!" — and the same last-modified date as ExamShieldSystem. Two parallel starts on the same problem, abandoned on the same day. The record does not say which one I meant to keep.
Two separate starts on the Green University Computer Club website, a month apart. Both still carry the unmodified create-next-app README, which is the honest signature of a project that got as far as npm and no further. The club site I actually maintained as Joint Information Secretary is not either of these.
The previous version of this portfolio. A single-page cybernetic machine that booted in standby, energised its conduits, spun up turbines and brought a rack online as you scrolled, with the projects as physical modules in that rack. It worked. It was also a machine that performed the same way for everyone who arrived, and that is the specific thing this version was built to stop doing. You are standing in its replacement.
A full booking client for Bangladesh Railway — train search, a visual seat map, multi-passenger booking up to four seats, ticket history, token auth, and a small Node proxy to get around the browser. It is finished in the way that matters and dead in the way that matters more: it is a front end for a system I do not control, so its lifetime was always going to be decided by somebody else's API.
Autonomous robot maze navigation, animated node by node: BFS, DFS and A* in a Python FastAPI backend, a Next.js front end, and — the part I would keep — queue, stack and priority-queue implementations written by hand instead of imported, because the whole point was to see the frontier expand. Not abandoned in disgrace. Just finished, and small.
An untouched create-next-app scaffold named "fun". Nothing was built. It is here because a graveyard that only contains dignified failures is a lie, and because the gap between naming an idea and starting it is the most common way work dies.
Unbuilt. These are open obligations, not achievements — every one of them is something that would make an existing claim stronger, and none of them has been done.
Epistemic Guard is a working instrument for measuring what AI assistance costs the person using it. The study it was built to run has not been run. No participants, no trials, no data.
Missing: Participants, ethics approval, and a protocol run. Until those exist there is no finding here, and the repository says so on its first screen.
CausalGuard reports 96.2% detection accuracy and 95.4% verification precision over 735 RCAEval failure cases. Those numbers come from its own benchmark harness, run by me.
Missing: A run by someone who did not write the harness. Until that happens the correct description of those figures is "self-reported", and that is how they are labelled everywhere on this site.
The Dry Bean scaling study retained 68 duplicate rows and never recomputed without them. It is listed in the study's own limitations section, generated automatically, which means the software flagged my omission rather than me remembering it.
Missing: A deduplicated re-run, and an honest note about whether it moves any of the reported conclusions.