Roselyn – RAG, LangChain, Python

An AI document-processing application for finance and accounting teams that were extracting and checking invoice and transaction data by hand, a slow and error-prone process. The application pulls the key financial fields from uploaded documents, validates them against predefined schemas and checks identity data. It stores the results as clean, structured records ready for review. Teams get a single interface for uploading documents and checking results.

• Built and deployed the end-to-end document-processing application, cutting manual processing time by 40%.
• Started from broad requirements: mapped the document flow, defined the minimum required fields and shipped a thin end-to-end version. Then tested it on real invoices and refined prompts and validation rules from stakeholder feedback.
• Designed the REST API contract between the React frontend and the FastAPI backend, with Pydantic schemas and consistent error responses.
• Engineered OpenAI extraction prompts and added schema-based validation and identity checks to keep the output reliable.
• Modelled and persisted financial records in PostgreSQL and containerized the service with Docker for deployment.
• Ran milestone-based client updates covering progress, next steps, blockers and decisions needed.

An automation solution for logistics and delivery operations teams that were moving data manually between order management and shipment tracking systems. Those teams were also triaging delivery exceptions and updating customers by hand. With the solution, shipment events trigger operational workflows automatically: dispatch updates, system-to-system sync and customer notifications. AI classifies each delivery exception and routes it to the right resolution path, so routine cases are handled without anyone stepping in.

• Built and launched event-driven workflows connecting order management and shipment tracking systems, automating routine dispatch and exception handling (metric needed).
• Designed AI-based delivery exception classification with OpenAI, routing each shipment issue to the right resolution path.
• Built conditional routing that triggers different actions by shipment status, exception type and operational rules.
• Kept both systems in sync two-way through REST APIs, webhooks, field mappings and data transformations.
• Automated customer notifications for key shipment and delivery events.
• Hardened reliability by debugging failed workflow runs, integration errors and API edge cases.

A loan-intake automation for lending teams whose loan officers spent much of their time re-keying online applications into the CRM, sorting supporting documents and chasing applicants for missing paperwork. The solution creates and updates CRM records from incoming applications and uses AI to classify submitted documents and check each application for completeness. It automatically emails applicants when something is missing. Loan officers get Slack alerts only when a human decision is needed.

• Designed and shipped the end-to-end loan intake workflow in n8n and OpenAI, removing manual data entry and applicant follow-ups for loan officers (metric needed).
• Expanded the original scope after talking to loan officers, targeting their real pain point: unnecessary manual follow-ups.
• Implemented AI document classification and completeness checks that decide the next step for each application automatically.
• Integrated the application portal with the CRM to create and update applicant and loan records automatically.
• Automated applicant follow-up emails through Gmail and Slack alerts for loan officers on status changes and exceptions.
• Added conditional routing and error handling, then tested and monitored the workflow in production.

A customer support chatbot for an education provider whose support team couldn’t keep up with prospective and current students’ questions. It answers questions about courses and enrollment from the provider’s own knowledge base, recommends courses based on the student’s needs and guides them through enrollment.

• Built and launched the RAG support chatbot, cutting average customer response time by 50%.
• Built the LangChain + OpenAI pipeline: ingestion, chunking, embeddings and vector-database retrieval over course and enrollment content.
• Combined vector and keyword retrieval and tuned chunking and ranking, so queries with exact course names and terms returned the right answers.
• Chose RAG over prompt-stuffing and fine-tuning to keep content easy to update, lower token costs and improve grounding.
• Designed conversational flows for course questions, personalized recommendations and enrollment help.