Sol Farahmand

Features include: PDF ingestion, parallel document review, learning-plan generation, human approval, quiz generation, expert evaluation and revision, RAG-grounded answer generation, multiple-choice generation, and finally student scoring and section-level performance.

A LangGraph-powered customer-operations demo that investigates support requests, retrieves customer and billing evidence, recommends a resolution with an LLM, and pauses high-value refunds for human approval.

The project demonstrates how to combine AI reasoning with deterministic business controls. The LLM classifies requests and recommends resolutions, while Python owns refund amounts, approval thresholds, routing, and execution safeguards.

What the project demonstrates:
• LangGraph state, nodes, conditional routing, interrupts, and checkpointing
• OpenAI structured outputs for predictable triage and resolution decisions
• Local customer and billing tools with typed return values
• Human approval before high-value financial actions
• Deterministic guardrails that prevent the LLM from choosing refund amounts or bypassing approval
• Separate recommendation, authorization, and execution stages

A support-ticket triage service built as a LangGraph workflow behind a FastAPI API. It classifies incoming tickets, assigns a priority, drafts a customer-facing response, evaluates that response against quality checks, and automatically revises it if it falls short.

A CrewAI Flow that audits a business process, finds bottlenecks, and redesigns it with a manager agent gating the findings, a human-in-the-loop approval step, and resumable state.

Give it a narrative description of a manual process (e.g., "how vendor invoices get approved"), and it will:
• Map the process into a numbered as-is step list and flag concrete bottlenecks
• Have a manager agent decide whether the findings are solid enough to act on, ambiguous enough to need a human, or not worth pursuing
• Pause for your approval when it's ambiguous
• Design a to-be process and a phased rollout plan
• Write the full audit to output/process_audit_.md
• Build as a learning project to demonstrate 5 CrewAI concepts together in 1 working system