EvoBot: An AI Mentor Bot for a 45-Day Capstone Course
3
Course Tools Exposed to the Agent
2
Swappable LLM Providers (Anthropic / OpenAI)
45-Day
Program Awareness, Tracked per Student
100%
Interactions Logged for Review
The Challenge
Students in a cohort course ask the same kinds of questions again and again — mentors spend their time on routine questions the course documents already answer, while questions that genuinely need a human get lost in the same channels.
Repetitive Questions
What's due this week, how a rubric item is weighted, what "done" means for a milestone asked over and over.
Mentor Time Drained
TAs spend time on questions the project brief and rubric already answer.
Escalations Get Lost
Grading disputes, extension requests and personal circumstances get buried in the same channels as routine questions.
No Auditable Handoff
Nothing reliably separates "an AI can answer this" from "a human must," or records that decision.
Our Solution
Built as a proof of concept, kept deliberately small: one process, one agent service, a file-based knowledge store and one log table.
Agent Built on the Claude Agent SDK
Each question runs through ClaudeSDKClient, with the bot's tools exposed as an in-process MCP server.
Resolve-vs-Escalate Policy
The system prompt decides what's answerable from course materials versus what needs a human mentor.
Swappable LLM Backend
One env var (LLM_PROVIDER) switches between Claude and OpenAI, with shared tool logic across both.
Grounded Retrieval
Search course materials keeps every direct answer traceable to the brief, rubric and milestones.search_course_materials keeps every direct answer traceable to the brief, rubric and milestones.
Business Impact
As a working proof of concept, EvoBot demonstrates both paths end to end: grounded direct answers, and a clean, auditable escalation to a human mentor.
Grounded Answers
Direct answers stay traceable to the project brief, rubric and milestones not freeform generation.
Clean Escalation Path
Grading disputes and edge cases reach a mentor queue with full context and a stated reason.
Zero Dropped Questions
Errors inside the agent loop are caught and logged — the student always gets a reply.
Full Visibility
Every turn is logged and reviewable live through the admin API.
Module Screenshots
A look inside the interfaces.
Evolvan's Role
Evolvan designed and built EvoBot end to end — the agent service, the provider-agnostic tool layer, and the logging and admin API that make its decisions auditable.
Agent & Tool Design
Evolvan built the three course tools — material search, progress lookup, and mentor escalation as an in-process MCP server.
Provider-Agnostic Architecture
Evolvan split the agent layer so Claude and OpenAI share one tool-logic module, switchable by a single env var.
Resolve/Escalate Policy
Evolvan authored the system-prompt policy that decides what the agent can answer versus what a mentor must.
Observability & Admin API
Evolvan built the SQLite logging and the FastAPI /health and /logs endpoints for live review.