Production systems I've built at Walmart Global Tech, and the things I build for fun.
Internal Walmart platforms — described here, but the code isn't public.
Walmart Global Tech
Asynchronous AI audit harness that auto-approves compliant seller listings and escalates edge cases.
Handles 200k+ daily seller listing updates. Enforces strict schema validation on LLM decisions, auto-approving high-probability compliant cases and routing ambiguous ones to human auditors.
Walmart Global Tech
Full-stack compliance platform plus a multi-agent RAG assistant over SOP documents, serving 60,000+ store associates.
Tracked 60+ critical metrics across 60,000+ store associates and prevented store-level compliance fines by grounding agent answers in internal SOP documentation.
Walmart Global Tech
End-to-end logistics investigation platform for analyzing anomalous delivery patterns.
Implemented OAuth2/OIDC RBAC, geospatial package search, and a WebGL-accelerated map UI on a React + Spring Boot + Elasticsearch stack.
Walmart Global Tech
Auditing tool for reviewing historical dispensing data under regulatory opioid monitoring.
A React and GraphQL tool that lets investigators streamline reviews and mitigate OPM penalties.
Walmart Global Tech
Distributed service logging multi-database CRUD operations with replay-by-key verification.
Built for data integrity verification: every write is captured to Kafka and replayable by key.
Public repositories, mostly in the RAG and AI-agent space. See also github.com/ranjan-apu.
A system design studio: diagram architectures, simulate real load, and practice interviews — all in the browser.
Draw a system on the canvas, push traffic through it, and watch queueing behavior emerge: latency percentiles climbing, queues filling, circuit breakers tripping. Then rehearse explaining it with guided interview packs.
Research harness for evaluating an LLM-powered intraday price-action trading agent.
Tests whether a structured LLM agent, given clean multi-timeframe price-action data, can generate useful BUY/SELL/SKIP/HOLD/EXIT signals in a walk-forward backtest. Research POC — it does not place broker orders.
A retrieval-augmented generation system that prioritizes retrieved document context over model priors.
LangChain orchestration over QdrantVectorStore with MistralAIEmbeddings and an OpenAI-compatible endpoint, falling back to general knowledge only when retrieval comes up empty.
A command-line lab demonstrating seven advanced RAG techniques for querying PDFs.
LangChain with Qdrant in Docker, built as a hands-on way to compare advanced retrieval strategies side by side.
A structured step-by-step code generation agent following a start → plan → action → observe loop.
Validates each step before execution, integrates pluggable tools (weather, system commands), and is designed for CLI or assistant integration.
Turns long-form videos into vertical shorts by extracting highlights with GPT-4 and Whisper.
Downloads a video, transcribes it with Whisper, identifies the most engaging segments with GPT-4, detects speakers, and crops them to vertical.
A minimal LangChain RAG system for chatting with PDF documents.
Fine-tunes DeepSeek R1 Distill 3B to predict 5-minute crypto candles from OHLCV data.
Pulls exchange data with CCXT, computes RSI/MACD/EMA with pandas_ta, then fine-tunes the model on normalized price and indicator sequences.
Competitive programming solutions, mostly C++.