Adwitiya Rahman

EaseMySemester, Exam revision product for first-year MBBS students

Live cohort with paying users. Next.js, TypeScript, PostgreSQL on Neon.

• –Built and shipped an invite-only revision product over 10 years of previous-year exam papers. Roughly 50 registered students in one cohort with no marketing, and 6 paying for the premium tier.
• –Designed the prediction engine as pure deterministic statistics rather than an LLM, chosen for auditability and run-to-run stability, because a ranking the student cannot interrogate is a ranking they will not act on. Measured 8 of 20 recommended questions appearing in an unseen exam paper, scored against a year the ranking had never seen.
• –Moved the paywall off the question bank and onto the prediction tier, on the argument that the bank is raw material while the ranking is the work already done, and people pay for the work already done.
• –Reversed my own information architecture after users showed it cost roughly 5 clicks to reach a question on exam night, cutting it to 0 to 3 without changing the data model.

Clarity, Finance ledger for a single-clinic practice

897 transaction records migrated out of sixteen months of spreadsheets, in daily use since. Next.js, TypeScript, Supabase, Gemini.

• –Shipped an LLM narration feature where typed code computes every figure and the model is architecturally forbidden from arithmetic, policed by a pure offline faithfulness gate that discards the entire narration if any numeral fails to trace back to its source.
• –Built a 100-case evaluation suite that runs nightly and fails the build on a faithfulness miss, with a separate tone judge deliberately kept non-blocking, because a subjective score is the wrong instrument to hold a release hostage to. Diagnosed a live faithfulness failure from the shape of the failing numbers, traced it to a prompt that forbade inventing figures but not deriving them, and took the suite from 86 of 100 to 100 of 100.
• –Retired the project's own headline metric after one user conversation revealed it was comparing two different jobs, and rebuilt the roadmap on what the user actually said rather than on the mechanism analysis.
• –Derived the product wedge from the data rather than from interviews: the source workbook was a 16.5-month cash receipt log with no patient entity, so it structurally could not answer who still owed money.

Semantic Search, Hybrid retrieval and RAG evaluation workbench

Internal tooling built for my own projects. The bench the other two are evaluated against. Python, FastAPI, SQLite, Ollama.

• –Built the measurement loop before tuning any retrieval setting: saved question sets replayed through any embedding, reranker and generation model combination, every stage timed and every run comparable.
• –Implemented hybrid retrieval, semantic embeddings fused with BM25 keyword search through weighted Reciprocal Rank Fusion, with an optional cross-encoder reranker and per-claim citations to exact book and page.
• –Root-caused a silent retrieval failure where a 4,096-token context window evicted retrieved passages mid-generation, a defect that presents as poor model quality and is not. Sequenced 5 concurrency defects by blast radius, deadlock first and the headline throughput fix last so the measurement could not mislead.

Wiki Pipeline, Self-writing knowledge base with an independent checking model

Runs hourly, unattended, with no review gate. Python, Claude, git, Windows Task Scheduler.

• –Specified an agent pipeline as two roles that do not report to each other: a writer that drafts each page from harvested sources, and a checker given those sources inline with no tools of its own, so it cannot go looking for agreement with the draft.
• –Ran it with no human review step on purpose, on the argument that a process needing me daily is a process I abandon, and put the safety in an independent checker, a deterministic linter and git as the undo button instead.
• –Treated hidden model thinking as a product constraint with a measured cost, after one uncapped checker spent roughly 21,000 tokens to make 3 edits while a zero-capped one approved a page in 4 seconds.
• –Built the redaction as a blocking gate rather than a promise: 0 of 204 harvested items reached a written page with anything the scan was built to catch.

Hexaware Technologies

Optum (UnitedHealth Group), Gurugram, India

Temenos, Bengaluru, India

Master of Business Administration

Indian Institute of Technology Kharagpur, Vinod Gupta School of Management

Specialisations in Information Technology and Marketing. Elected student representative.

B.Tech, Computer Science and Engineering

SRM Institute of Science and Technology

Elected Vice-President and then President of CHiPSET, one of the most active technical societies on campus.