Case studies
The problem, then the build.
Four projects in enough detail to judge whether I'd be useful on yours — what the actual constraint was, what got built, and where it ended up.
AI Knowledge Base & RAG Systems
AI engineering · Python
The hard part of a document Q&A system isn't the model — it's retrieval. A generic chatbot bolted onto a document store answers confidently from nothing, and loses the room the first time someone checks a citation.
The problem
Organisations sit on contracts, manuals, support histories and internal documentation that only a long-serving employee can navigate. The knowledge exists; it just isn't reachable by anyone who needs it at the moment they need it.
What I built
Ingestion and chunking tuned to how the source documents are actually structured, embeddings and vector search — Pinecone, Qdrant, FAISS or pgvector depending on the deployment — retrieval with deliberate context-window management, and a FastAPI service layer in front of the LLM APIs.
How it holds up
Answers stay tied to the passages they came from, so anyone can check the source. Retrieval is the layer that gets tuned over time; the model behind it stays swappable. Where data can't leave the building, the same architecture runs against local models via Ollama.
Sparkle & Co
E-commerce · sole technical owner · 8+ years
A growing nail-products retailer with no in-house technical team. For eight-plus years I was the entire technical function — storefront, server, integrations, reporting and search.
The problem
Every technical need landed with one person: a new storefront feature, a server falling over on a Sunday, a reporting question from the owner, an SEO decision, and eventually a full platform migration. No team to escalate to.
What I built
The WooCommerce storefront and the custom plugins and internal tools around it, sales and inventory reporting, server administration and deployment, continuous SEO work, a later migration to Shopify, and an AI-based nail simulator so customers could preview products.
How it holds up
Eight-plus years of unbroken ownership across two platforms — the kind of engagement where you inherit every decision you made three years earlier and have to live with it. Relationships like this are the closest thing to a real reference a freelancer has.
Judtho
Product engineering · web + mobile
A chess variant with pieces standard engines don't have, and 2-, 3- and 4-player formats that standard turn logic doesn't handle. Nothing off the shelf could be adapted, so the rules engine was written from scratch.
The problem
A chess variant lives or dies on whether the rules engine is correct. This one adds pieces with movement patterns existing chess libraries don't implement, and supports three- and four-player games where turn order and elimination don't behave like standard chess.
What I built
Move generation and validation written from first principles, multi-player turn and elimination handling, a React web client for play, Python services behind it, and a Flutter path for mobile delivery.
How it holds up
Live and playable at judtho.com — which makes it one of the few things in any portfolio you can fully evaluate without taking the author's word for anything.
Audio AI — transcription & music tooling
Python · AI · signal processing
A run of tools built around the same frustration: audio work that is technically solvable but painfully manual, and Indian-language audio that mainstream tools handle badly.
The problem
Transcription in Hindi and Gujarati is poorly served by mainstream tools. Meanwhile routine studio tasks — pitch correction, repeated effect chains, stripping vocals to recover an instrumental — consume hours of identical manual work.
What I built
A multilingual Hindi/Gujarati/English transcription app on faster-whisper with a React and FastAPI core, wrapped as an Electron desktop app with live microphone recording and subtitle export. A song-to-instrumental converter using Demucs for stem separation, Basic Pitch for melody extraction and FluidSynth for rendering. And a Python autotune engine handling pitch correction plus repeatable compression, echo and reverb chains.
How it holds up
Working desktop and web tools that replaced processes previously done by hand. The interesting engineering is in the pipeline design — these are long-running, resource-hungry jobs that have to stay responsive to a user sitting in front of them.