Karan Kumar
Full Stack Software Engineer — building production systems and, right now, retrieval-augmented generation from scratch.
About me
I build the layer between AI and production — where models meet real users, real data, and real failure modes. I built the backend for Brarista, an AI fitting tool for Shopify storefronts, and I'm now building a RAG pipeline from scratch — Ollama, ChromaDB, sentence-aware chunking — because I'd rather understand retrieval than call an API.
I ship, then stick around to make sure it holds.
Location
FLORENCE_IT
Focus
BACKEND_LLM
Education
MSc, Univ. of Florence
Status
OPERATIONAL
Experience
- Dec 2022 — Jan 2025
Full Stack AI Engineer
Concept Recall
Worked across the full stack on Brarista, an AI-driven bra-fitting platform for e-commerce clients, and stepped up to own backend delivery (Python, AWS Lambda, API Gateway) during a team shortage, learning the infrastructure on the job.
Diagnosed and fixed a multi-tenant response-routing bug in Brarista's backend that was silently sending users to a broken "email your results" fallback instead of their actual size recommendation; traced the issue across frontend and backend and corrected it without a full rewrite of the surrounding system.
Packaged the Brarista fitting tool as a reusable React component library, adopted by two internal teams for additional Shopify storefronts, and contributed to CI/CD automation, Cypress E2E test coverage, and onboarding new engineers through PR reviews.
- May 2021 — Dec 2022
Freelance Full Stack Developer
Fiverr (Level 1 Seller)
- Delivered 30+ projects for 15+ clients spanning REST API integrations, authentication systems, backend automation, and responsive React/Next.js frontends with custom animations.
- Debugged cross-browser JavaScript and CSS issues for long-term clients.
- Maintained a consistent 5-star client rating throughout.
Projects
Brarista AI Fitting Tool
An AI-powered size-calculation backend for an e-commerce bra-fitting platform, serving real-time predictions to live Shopify storefronts.
A Next.js portfolio with a self-built content system underneath it — draft/publish revisions, a section-based page composer, role-based admin access, and an admin-triggered JSON import pipeline for content — running on Vercel against a managed Postgres database.
A production-grade real-time chat application built as six independent Go microservices. The architecture tackles WebSocket fan-out across instances via Redis Streams, JWT token rotation with race-condition safety, a three-layer blocking cache, and a raw-TCP WebSocket tunnel at the API gateway.
Skills
Education
- Sep 2024 — Aug 2026
Master's in Software: Science & Technology
University of Florence (UNIFI) — Firenze, IT
- Erasmus Nazionale exchange (6 months) at the University of Camerino (UNICAM), with coursework in software testing, multi-agent systems, and process mining.
- Coursework in Distributed Systems Design and concurrent programming with Go, which motivated building GoChat independently.
- Oct 2018 — Oct 2022
BS in Information Technology
Quaid-e-Awam University of Engineering, Science & Technology
- CGPA 3.8/4.0, ranked 4th in graduating cohort.
- Technical Team Member for TechnoMind-2k19, organizing 14+ technical competitions.
Building Now
A local-first RAG pipeline from scratch — Ollama for inference, ChromaDB for vector storage, sentence-aware chunking, and LLM-based query routing across multiple documents, with a Streamlit interface on top.