Selected engineering work

Production engineering at VdoCipher, alongside personal projects exploring document retrieval, AI interaction and everyday product problems.

Professional work · 2019–2025

VdoCipher: Faster search and frontend builds

Problem
Customers needed faster search, while developers needed shorter feedback loops when changing the frontend.
My work
Built distributed search with PostgreSQL and OpenSearch, and moved the frontend build from Webpack to Vite.
Result
Search response time fell from 2–3 seconds to about 200 ms, and frontend build times fell by about 90%.

React · TypeScript · PostgreSQL · OpenSearch · Vite

Personal project · Document Q&A

AskBase: Answers you can trace to their sources

Problem
Document answers need to be useful and verifiable. A fluent response alone does not show where the information came from.
My work
Built a PDF and URL ingestion pipeline: extract content, split it into chunks, embed it, and store it in pgvector. Added streamed answers, clickable source citations, feedback-based chunk quality scores, and an admin dashboard.
Result
The MVP supports document upload, source-linked answers, and monitoring of feedback and retrieval gaps.
Recorded demonstration from the project repository. Try the live demo below.

Next.js · TypeScript · Supabase PostgreSQL · pgvector · OpenAI

Personal project · UAE outings

Somewhere: A good day out, without all the searching.

Problem
I kept coming across places to explore around the UAE, then forgetting them when we actually had a day off.
My work
So I built Somewhere: a guide to save those places and help decide where to go based on your emirate, available time and mood.
Result
A mobile-friendly guide with saved places, personal notes and a day planner. The picker suggests up to three outings and explains each match.

React · Vite · PWA

Personal prototype · AI group chat

Mesh: An AI participant that knows when to stay quiet

Problem
An AI that answers every message interrupts a group conversation. Useful participation requires deciding when a response is warranted.
My work
Built a two-stage flow: a lightweight relevance classifier decides whether to respond, then a separate model streams the answer. A context panel extracts decisions, tasks, budget items, and links, with references back to source messages.
Result
The prototype demonstrates selective AI replies, streaming, room-level settings, and a guided group conversation.

Next.js · TypeScript · Vercel AI SDK · OpenAI · SSE