01
APPROVED_INFORMATION
Answers start with approved business information, so the team can check where a response came from.
01 · Semai AI Support · Agentic AI / RAG
Support teams: answers from approved information, with human help
Challenge: Semai needed customer support that could use approved information without handing control to a black-box chatbot. Solution: I designed the AI workflow, API, search, tool, handoff, and operator layers with Python, FastAPI, LangGraph, LangChain, and React. Outcome: a controlled support design with human handoff and operator feedback; no live performance metric is claimed. Next: Message me to map this to your support workflow.
AI Product Architect & Lead Engineer2026Built
Case study
What was built and why it mattered.
Outcome, ownership, architecture, and evidence in one view.
Semai turns approved business information into a support tool that can answer routine customer questions while keeping a clear path to a person. The goal was useful automation that the team could review and improve.
I designed the AI workflow, API, and support screens: search for approved answers, limited tool use, clear handoffs for sensitive or unresolved issues, and a dashboard for reviewing conversations and updating the knowledge base.
The implementation combines Python, FastAPI, LangGraph/LangChain patterns, OpenAI-compatible models, and a React operator experience. It shows how AI support can fit into a team's daily work without hiding the important decisions.
The support flow connects approved information and limited tools, while the team reviews handoffs and improves the source material.Solid = primary flow · dashed = support / control path
Download editable Excalidraw02
CONTROLLED_WORKFLOW
LangGraph-style workflow control connects the request, search, tool use, response, and handoff steps.
03
HUMAN_HANDOFF
Handoff records and operator review keep people responsible for complaints, unusual cases, and decisions that need judgement.
04
SUPPORT_TEAM_TOOLS
Conversation, knowledge, and reporting screens help the support team maintain the service.
The brief
Semai was conceived as an AI support product for businesses that needed faster, more consistent customer assistance without surrendering control to an opaque chatbot. The real requirement was an operating layer: one that uses approved knowledge, captures context, knows when to stop, and gives a human team a practical way to improve it.
What I owned
I designed the product architecture across the agent, API, knowledge, and operator layers. That included the conversation lifecycle, retrieval and tool boundaries, escalation behavior, and the surfaces that let a business review conversations and maintain its knowledge base.
Architecture
Loading diagram…
Engineering decisions that matter
Grounded answers over fluent guesses
The system is structured around approved business knowledge and retrieval. The objective is not simply to make the assistant sound confident; it is to make its answers useful, bounded, and maintainable as business information changes.
Escalation is part of the product, not a fallback
Complaints, sensitive situations, and unresolved questions are captured as explicit escalation records with priority and context. This means the team can take ownership of the work instead of treating an unhelpful AI answer as the end of the interaction.
Operators need a feedback loop
The companion dashboard is designed for conversation review, knowledge maintenance, and analytics. An AI support system improves when operators can see where customers get stuck and refine the material that shapes future responses.
Implementation
- Python services and FastAPI for the API and integration boundary.
- LangGraph/LangChain patterns for agent orchestration and tool-aware workflows.
- OpenAI-compatible model access and retrieval-backed knowledge handling.
- React-based web and dashboard surfaces for customer interaction and operations.
- Structured escalation tooling for complaints and human follow-up.
What this case study proves
Semai shows how I approach production AI as a system-design problem: connect models to trustworthy context, give people control of exceptions, and build the operational feedback loop that keeps the product useful after launch.