Karamvir Singh

Developed an end-to-end computer vision application following the City of Ottawa policies to guide waste classification. Users can click a picture of the waste item and get guided towards the right bin color.

Automated a trading bot based on strong price move detection using GenAI multi-modal capabilities. The bot detects tradable strong moves (inspired by SMC-based concepts like MSS and BOS) with 40% more accuracy than standard methods like ATR.

I built a RAG pipeline using Amazon Bedrock for software projects (code and documentation) for a development shop, using AWS Knowledge Base and AWS OpenSearch for indexing and vector databases. I integrated, retrieved, and generated an API with a Python front end, secured with IAM.

• Built the core back-end foundation for a hyper-personalized eCommerce platform.
• Architected the end-to-end retrieval and ranking stack, combining vector search (Pinecone), hybrid metadata filtering, and LLM-based query understanding to convert natural language input into structured intents and high-relevance results.
• Focused on speed and accuracy, targeting 0.7+ f1 scores and sub 500 ms responses.
• Developed scalable FastAPI services and async pipelines for search, recommendations, and user memory, with integrated caching (Redis), rate limiting, retries, and fallbacks to ensure low latency and resilience during rapid product iteration.
• Designed and implemented the personalization and data-signal layer, capturing engagement events—including searches and purchases—and feeding these into downstream ranking systems to continuously improve relevance.
• Worked closely with research on retrieval evaluation and with product/design teams to refine user-facing experiences, enabling fast experimentation and reliable system performance across the launch.