Charuka Gunawardhane

Approach

I designed a modular machine learning forecasting system that separates demand into baseline, promotional uplift, and uncertainty buffers. Key decisions: - Use XGBoost for its ability to model non-linear demand patterns and handle mixed feature types - Build a single global model across item × outlet combinations to improve generalization - Introduce multi-horizon forecasting aligned with supply chain decision timelines - Replace static safety stock rules with error-driven buffer calculations - Automate the full pipeline from data ingestion to order management system integration

Approach

We designed and delivered a hierarchical agentic AI system orchestrated using LangGraph, integrated with a custom ecommerce backend. A central supervisor agent routes user intent to specialised sub-agents responsible for product search, recipe intelligence, cart management, and order execution. Product retrieval combines lexical search (pg_trgm) with semantic vector search (pgvector) to handle misspellings, synonyms, and local grocery terminology, with business-aware re-ranking applied on top. A draft-and-confirm transaction model ensures no order is placed, modified, or cancelled without explicit user approval. Conversation state is persisted in PostgreSQL and rehydrated on each LLM turn, enabling stateful multi-turn interactions without relying on LLM memory. Multimodal inputs — text lists, conversational queries, and OCR-extracted image lists — are unified through a single downstream processing pipeline.

Approach

We designed and implemented a two-pipeline RAG architecture on AWS. An offline batch ingestion pipeline (ECS Fargate) pulls HR documents from S3 and SharePoint, extracts and cleans text, chunks documents into 200–300 token windows with overlap, generates 1536-dimensional embeddings via Amazon Titan Embeddings v2, and bulk-indexes chunks with metadata (business unit, country, ACL groups) into Amazon OpenSearch's k-NN Vector Engine. A real-time query pipeline handles each employee request: the authenticated query passes through API Gateway and AWS WAF to an AWS Lambda orchestrator, which embeds the question with Titan, runs a k-NN search on OpenSearch with RBAC metadata filters to restrict results to the user's business unit and country, constructs a grounded prompt from the top-K retrieved chunks, and calls Claude 3.7 Sonnet on Amazon Bedrock for answer generation. The two pipelines share the same OpenSearch index cluster, keeping infrastructure lean while keeping ingestion fully decoupled from query-time logic. The implementation (GitHub) uses FastAPI, GPT-4o, and Qdrant as a portable local stack that mirrors the production AWS design.