aug 2026 – present
td synnexagent
what it is
a support assistant for microsoft azure questions, built with a team as a td synnex-sponsored capstone at clemson. it answers from trusted microsoft documentation and shows where each answer came from. when the docs don't support an answer, it says so instead of guessing.
- ask
- search
- answer or escalate
what i built
- cited answers: the first frontend for the assistant, with a sources panel and citations linked to the exact documentation files.
- hosting: deployment to azure container apps with microsoft sign-in.
- escalation: connecting the model’s decisions to the live pipeline, so it can produce all five escalation reasons.
- clarifying questions: when a request is ambiguous, the assistant asks instead of picking a product for the user.
- conversations: multi-turn threads stored in shared blob storage.
- evaluation: i lead a four-person team on this. i built a 152-item test set with a separate held-out split, plus prompt-injection tests, including instructions planted inside retrieved documents.
when it can’t answer
The documentation found doesn't clearly support an answer to this question.
why it matters
a confident wrong answer is worse than no answer in support work. so the agent is built to know when to stop.
where it stands
- running
- the assistant is deployed on azure, and we test it against live azure services.
- results
- we’ve run benchmarks, and the results are still being reviewed, so i’m not sharing numbers yet.
- planned
- comparing it with web search and with a language model on its own, across accuracy, citation quality, latency and token usage.