Node.js · React Native · PostgreSQL · LLM systems
I design distributed backend systems — event-driven queues, GraphQL APIs, vector retrieval and production LLM pipelines — and the React Native clients that consume them.
Mulltiply · B2B commerce platform for Indian retail
Mulltiply
Ordered by depth. Everything listed here is something I've shipped with.
Live products and personal systems — not coursework.
The retailer-facing React Native app, serving a catalogue of 10,000+ SKUs with real-time stock, GST invoicing, multilingual UI, deep linking, RBAC and OTA updates. I'm the second-highest contributor across its history — 1,023 commits, 16 screens, 12 components and the full analytics layer.
~40% faster cold start through lazy-loaded home surfaces, shimmer placeholders and render profiling.
A multi-tenant RAG platform where an organisation uploads its internal documents (HR, policies, SOPs) and queries them in natural language, with versioning and role-based access.
PDF parsing → chunking → 1536-dim embeddings in pgvector with an ivfflat cosine index; tenant isolation enforced at the schema level by org-scoped foreign keys; and a hand-rolled Postgres job queue driving the embedding worker rather than an off-the-shelf broker.
A unified engineering dashboard that harvests your GitHub history — including private organisation repositories — and synthesises it into a career and impact report.
Walks the GitHub GraphQL API with author-filtered commit pagination, honours rate-limit hints with exponential backoff, and harvests repositories through a fixed-size concurrency pool so a two-year scan survives throttling. Turborepo monorepo with detached background scans the client polls.
Six systems I own or built from scratch. All production, all at Mulltiply.
An LLM query parser that converts natural-language Hindi, Hinglish and Devanagari product queries into a structured Meilisearch query plus a filter expression — transliterating चना दाल → chana dal and detecting colloquial price intent like "500 tak", "1000 ke aaspas", "kitne ka".
Treated as a compiler, not a chatbot: price cues resolve to concrete numeric bounds at parse time, approximate prices expand to ±10% ranges, and a bare number without a price cue is explicitly not read as a price — so the model can't invent filters.
Owned the connector layer that syncs the platform with the offline accounting software Indian distributors actually run — Marg, Tally, Busy and Zoho, plus Shopify.
Built Zoho's full OAuth flow, Marg item/pricing/order/ledger/invoice sync with batch-wise processing for oversized payloads, and Busy CRUD for items, customers, invoices and stock. Wrote the transform layer that reconciles four incompatible ERP schemas into one internal model.
Retailers order from paper invoices and handwritten lists. This turns a photo into structured order lines.
Multi-file upload and parsing, an LLM pass that structures raw OCR text, a preview step before anything is committed, and an ocrTrainingData module I created to capture results for model improvement. Contributed to the Python service behind it.
A zero-typing order path for users with low literacy or low comfort with apps.
Speech → transcript → LLM parser, with automatic language detection, wired through to the cart. Removes the keyboard from the critical path entirely.
"Similar products" without a hand-built taxonomy.
Added embedding columns to the items table (SKU attributes and AI-generated tags included in the embedding text), fixed vector dimensionality, and built a similarity computation worker plus cron-driven embedding jobs. Created the recommendation module from scratch.
Notification triggers were hardcoded — every change to who gets told what needed a deploy.
Replaced them with a rule-based engine: configurable event keys, recipient types (employee, custom, retailer, salesman) with validation and normalisation, and a default event registry. Ops now self-serve. Also traced and fixed a duplicate-notification bug caused by salesman details leaking into workflow events.
Longer pieces on retrieval, recommendations and event-driven architecture. Read all →
Parsing Hinglish product search, where "kids shoes 7 number 400gm 500 rs tak" contains four numbers and exactly one of them is money.
Why a pure embedding recommender returns beautiful, useless results — and how blending it with co-occurrence, category and attribute signals fixes that.
Replacing hardcoded notification triggers with a rules engine, and the duplicate-message bug that only appeared when one person held two roles.
Raj Kumar Goel Institute of Technology · Ghaziabad, UP
Coursework in cloud computing, distributed and IoT system architecture, sensor networks and network planning.
Fastest by email. Based in Ghaziabad, working in Noida, Uttar Pradesh.