Nik Patel

Mood-aware Bhagavad Gita reading companion (React)

A mobile-first companion that surfaces Gita verses based on the reader's emotional state, in English and Hindi.

Built for daily study: users pick a mood — anxious, grateful, angry, grieving, peaceful, confused — or describe how they're feeling in their own words, and the app surfaces five to six verses curated to that state. Every verse renders Sanskrit shloka, transliteration, and a canonical translation. The mood-to-verse mapping is hand-built and grounded in traditional commentary, not generated by an algorithm. For a sacred text, that curation decision IS the product decision.

Engineering choices in service of the use case: fully offline after initial load — all 700 verses bundled, zero network calls, instant English / Hindi switch. Light / dark / system theming with proper OS-preference detection and persistence. Mobile-first responsive layout tuned for one-handed phone reading, with read-aloud, share, and bookmark on every verse. Honest fallbacks throughout: when a free-form mood input doesn't match a known feeling, the app tells the user and asks them to rephrase, instead of silently picking arbitrary verses.

Shipped solo. English and Hindi only — Gujarati and other Indian languages were considered but dropped: open-source verse data is thin, and I'd rather ship something honest than something half-translated.

Tech: React, TypeScript, Tailwind, Vite

Live demo → · GitHub →

Browser-based RAG chat for a B2B documentation knowledge base

A retrieval-augmented chat interface running entirely in the browser, indexing a client's full product documentation and answering questions with cited sources.

Designed and built a browser-based RAG (retrieval-augmented generation) chat for a B2B SaaS client's customer-facing documentation. Used Google's text-embedding-004 for embeddings and Claude Sonnet for response generation. The interesting parts were the operational ones: most "RAG demos" assume direct API access from the browser — production environments rarely allow that. I designed a CORS proxy fallback strategy plus a manual paste workflow so the system degrades gracefully when corporate networks block direct API calls.

Built so the client could ship it without standing up a Python backend or vector DB infrastructure they'd then have to maintain.

Tech: JavaScript, Google text-embedding-004, Claude Sonnet, CORS proxy fallback

Architecture walkthrough → (anonymized — happy to walk through architecture on a call)

Real-time inventory search for real estate agents

A Rails + React platform serving live pricing and availability across millions of property listings, with sub-second filtering on 15+ dimensions.

Built for a real estate SaaS company whose agents needed to generate up-to-the-minute, client-ready property lists in seconds, not minutes. The platform indexes millions of inventory units across multiple cities, with 15+ filter dimensions (price, layout, amenities, proximity, availability windows, builder, possession date, and more) returning filtered results in milliseconds.

The hard parts: keeping the inventory data fresh as feeds arrived from dozens of sources with conflicting schemas, and making complex multi-filter queries feel instant. Used PostgreSQL with carefully tuned indexes and Redis for hot-path caching, plus a write-side normalization layer that absorbed the source-data chaos so the read side stayed clean.

Tech: Rails, PostgreSQL, React, Redis

Architecture walkthrough → (anonymized — full architecture available on call)