I built hybrid retrieval, then measured it. It tied.
A 27-question eval where hybrid search tied with plain vector search, and why the test set was the problem.
I build full-stack andapplied AI systemsFrom ideato production.
Final-year BCA at Manipal University Jaipur, on the Kalvium software product engineering track. Right now I'm interning in the IT department of a Delhi government power utility. Looking for an SDE or GenAI internship from September 2026.

hello
Hi, I'm Aditi Vashishtha.
I build alone, which means I break things alone too. That's mostly how I've learned anything.
how I show work
Every project says what it doesn't do.
The limitation sits on the card next to the stack, because that is the real state of the thing. A page that only lists what works is describing a good day.
FOR EXAMPLEa11y Agent
73 to 76% is the share of axe-detected violations of the four targeted types. It is not a claim about the whole page.
FOR EXAMPLEFamiliar
Built as a client-rendered SPA, so shared article links preview blank and Google struggles to index it. Wrong architecture for a publishing platform.
shipped
how I ship

An accessibility agent that isn't allowed to grade its own homework.
Scans a live page, proposes a context-aware fix for one of four issue types, applies it, re-scans to confirm the violation actually cleared, and retries once with the axe error as feedback before failing honestly.
The model only makes judgment calls. axe-core owns detection and verification. Deterministic contrast fixes sit in the pipeline as a control case, and they caught two real bugs I would otherwise have shipped.

Document Q&A that has to survive two retrievers before it answers.
Postgres full-text search and pgvector run side by side, and reciprocal rank fusion merges the two rankings. A passage has to place well on keyword match and on semantic similarity before it gets cited.
The unglamorous half: PDFs are byte-validated before parsing, requests are rate limited per IP, there's a global daily cost cap so a bad day can't become a bad bill, and uploads delete themselves after thirty minutes.

A publishing platform with the boring parts done properly.
Rich-text editor, infinite feed, threaded comments, likes, bookmarks, follows and in-app notifications. Frontend, backend and deployment, all mine.
The auth surface is the part I'd point at: JWT sessions, Google OAuth with account linking, and single-use expiring tokens for email verification and password reset. Notifications are isolated on purpose, so one failing never breaks the like or comment that triggered it.

Cafe discovery for Jaipur, ranked so new places can't game it.
44 Jaipur cafes drawn from a public Zomato dataset, with natural language search. GPT-4o-mini parses the query, Gemini takes over when it doesn't.
Ranking is a Bayesian weighted score rather than a raw average, so a cafe with four five-star reviews doesn't outrank one with two hundred good ones. Custom pixel-art interface, built and deployed solo.

The official site for an international academic conference.
Multi-page site for a conference at Manipal University Jaipur, with scroll-triggered animations and responsive carousels, serving attendees and organisers.
I built hybrid retrieval, then measured it. It tied.
A 27-question eval where hybrid search tied with plain vector search, and why the test set was the problem.