AI & LLM Feature Integration

Adding an LLM call to a product takes an afternoon. Making it fast, accurate and affordable enough for real users takes engineering: latency budgets, retries, cost control, and a plan for when the model gets it wrong.

At MUST Company I developed MustBreak, an AI-powered real-time meeting platform for transcription, translation and multilingual collaboration across distributed teams. It produces AI-generated summaries, keeps conversation history searchable, and connects to Slack workflows, all on low-latency real-time infrastructure built with Python, FastAPI and DynamoDB.

StartDescribe brings the same thinking to e-commerce. Its AI pipeline generates product descriptions and images with OpenAI and Perplexity, inside a multi-tenant SaaS with Stripe subscriptions. Optimised generation workflows cut content generation time by 30%.

For products that need to answer from their own data, I build retrieval-augmented generation (RAG) with pgvector or Pinecone, and agentic workflows with LangGraph, on OpenAI and Anthropic APIs. Where accuracy matters, the design includes a review step so a person can check the output before it reaches a customer.

I'm an AWS Certified AI Practitioner, and I treat AI features like any other production system: measured, tested against real inputs, and easy to roll back.