Skip to content
How much to read
Adwitiya Rahman
EmailResume
All work

Every project, one page each

The same five questions answered about every one of them: what was broken, who for, what I decided, what it measured, and what I refused to build. So they can be compared without reading fiveessays. Each links to the full version. Print this page and you have the whole portfolio as a single PDF, with the caveats still attached to the numbers.

Brief · 2026 · Product, research and engineering

EaseMySemester

A revision tool for first-year medical students, now on its second version. It narrows to the hour before the paper, answers every question out of nine textbooks with the page number attached, and charges nobody.

Problem
A question bank of past questions, reached on exam night with an hour left and no time to find the answers.
User
First-year MBBS students, one batch of roughly fifty.
Decision
Build for the hour before the paper, not the term. Exam night won, and most of v1's structure went with it.
Number
809 questions answered out of nine textbooks, each citing its book and pages. 82, or 10.1%, decline rather than guess, nearly all multi-part questions with one unsourced sub-part. Corpus or retrieval, I have not separated which.
Refused
No public signup, no ingestion pipeline, no paid tier. Loading PDFs stays an offline job I run myself.
  • 34 of ~50students in the batch moved to v2, with no marketing
  • 809questions answered from nine books, cited to the page

The full case study is about 10 minutes on how it was decided, measured, and where it can still fool me.

Brief · 2026 · Product, research and engineering

Clarity

A bespoke finance ledger for a dental clinic, replacing a 16-month Excel workflow. 897 records came across from the spreadsheet, and it is now being rebuilt around a second user who must never see a rupee figure.

Problem
Sixteen months of clinic finances in a cash receipt log with no patient in it, so it structurally could not answer who still owed money.
User
A dentist I know, running one small clinic, entering transactions between patients. The rebuild adds his attendant, who must never see a figure.
Decision
The model narrates and typed code computes. A numeral it cannot trace back to a source does not reach the screen.
Number
100 of 100 eval cases pass, and the build fails if they stop. The first run scored 86 and caught a real defect. The tone judge is deliberately left non-blocking.
Refused
The tone scorer never gates a release, the payment QR automation stays unbuilt, and the keyboard shortcuts he told me he does not use are coming back out.
  • 100-caseeval suite, fails the build
  • 86/100first run, caught a real defect

The full case study is about 10 minutes on how it was decided, measured, and where it can still fool me.

Brief · 2026 · Product, specification and review

Wiki Pipeline

A knowledge base that rewrites itself from my own repos every night, checked by a second model that has no tools and cannot go looking for agreement, plus a desktop app I use to drive the whole thing.

Problem
A hand-written knowledge base behind this site that went stale the day the projects moved on, and that I would not keep up by hand.
User
One user, and that user is me. It is the caveat on every number here.
Decision
Let it publish without me. A process that needs me every day is one I abandon, and an abandoned process still looks like it is working.
Number
200 pages written, and every one of them checked by a model that did not write it. One user, running hourly since 14 September 2026. Separately, 0 of 204 harvested items got past the redaction scan.
Refused
No remote on the repository and no hosting for the vault, and no attempt to have it write anything a reader sees. I write those.
  • 200 pageseach checked by a model that did not write it
  • 0 of 204harvested items leaked past the redaction scan

The full case study is about 10 minutes on how it was decided, measured, and where it can still fool me.

Brief · 2026 · Product and engineering

Semantic Search

A local, offline hybrid search and RAG evaluation workbench over PDF textbooks: semantic plus BM25, fused with Reciprocal Rank Fusion, every run benchmarked and comparable.

Problem
No way to tell whether a retrieval change helped, which makes every setting a thing to argue about rather than a thing to compare.
User
One user, and that user is me. It is the bench the other two projects get judged on.
Decision
Measure first. The benchmark runner existed before a single retrieval setting had been tuned.
Number
5 concurrency defects fixed in blast-radius order, deadlock first. The headline throughput fix went last on purpose, so the measurement could not flatter the fixes underneath it.
Refused
No sweeps across model combinations, no answer-quality scoring yet, and a locking issue filed rather than rushed into a change that had no room for it.
  • 5 defectsfixed in blast-radius order
  • 0 GPUsCPU-only, embeddings included

No demo for this one. It runs on my own hardware with my own books on it, and it is the bench I test the other two projects against rather than something built to be handed out.

The full case study is about 6 minutes on how it was decided, measured, and where it can still fool me.

Brief · 2026 · Product, research and engineering

EaseMySemester v1

The first version: ten years of scattered exam papers turned into a tagged, browsable question bank, with a ranking that guessed what would be asked next. Six students paid for that ranking. It is still running, and v2 replaced it.

Problem
Ten years of past exam papers scattered across drives and group chats, with no way to see what had actually been asked repeatedly.
User
First-year MBBS students, roughly fifty in one batch, invite only and never marketed.
Decision
No model in the ranking. The question was countable, so it was counted, and the counting was graded against a held-out year before it shipped.
Number
8 of 20 predicted questions turned up in the real paper. One cycle, one cohort, scored against a 2026 paper the ranking had never seen. A signal, not a validated model, and the question-level control against that same paper is still unrun.
Refused
No LLM in the ranking, no public signup, and no ingestion pipeline: getting PDFs into the database stayed an offline job I ran myself.
  • 40%strike rate on predicted questions
  • 6 paidstudents, pay what you can, before the paywall came out

The full case study is about 7 minutes on how it was decided, measured, and where it can still fool me.

The full case studies: the evidence behind each of these, and the parts that did not work.