Danyal Ali Asghar.

I built this platform to reduce manual lease review for life sciences REITs. I combined PDFPlumber, Azure Mistral OCR and dynamic heading extraction to handle varied biotech and pharmaceutical lease formats, then used OpenAI and Azure Mistral to map the content into organized fields and metadata.

I also built the processing around the documents: tenant isolation in Azure Blob Storage and Cosmos DB, validation and graceful handling of non-critical failures. I added streaming uploads, progress reporting and stage-level timing so the team could follow a document through processing.

I built a retrieval-augmented knowledge assistant with OpenAI, LangChain, LlamaIndex, Pinecone and FastAPI. I combined embedding search with prompt chains so questions could draw on relevant material from internal documentation and policy archives.

I connected retrieval and response generation through an API, bringing document search into a question-and-answer workflow. I focused the project on a practical use of generative AI: helping people work with information already held in their documents.

I connected Azure OpenAI, Azure Functions, Python, Airflow and dbt to orchestrate model inference alongside data transformations. I brought these steps into scheduled workflows that produced insights and alerts.

I included quality checks in the workflow so validation sat alongside the processing work. I approached AI inference as one part of a data pipeline, with orchestration and transformations supporting the final output.

At dinCloud, I developed a voice-assistant prototype that connected Whisper speech-to-text with GPT-4 responses. I built the path from spoken input to a conversational answer, bringing speech recognition and language generation into one workflow.

I refined prompt templates to improve conversational response quality and latency. I used the prototype to work through those two practical concerns together: how the assistant responded and how long someone had to wait for its answer.