Internship Case Study · Metro South Chamber of Commerce
Designing and building two AI-powered assistants to improve internal workflows and public information access.
During my internship at the Metro South Chamber of Commerce, I built two AI chatbot systems: an internal AI assistant for Chamber staff over accreditation documents, and a public-facing website chatbot for visitors and members. Both used a retrieval-grounded (RAG) approach with gpt-4o-mini, strict grounding, and fallback handling to prevent hallucinations. A major component was processing newsletter PDFs and website content into a searchable knowledge base. The project was presented at the Student Arts & Research Symposium (StARS) at Bridgewater State University with a live demo.
Overview
Organization
Metro South Chamber of Commerce
Role
AI Chatbot Developer Intern
Duration
Jan 2026 – May 2026
Systems Built
2 AI Chatbot Systems
Focus Areas
RAG · gpt-4o-mini · FastAPI · Prompt Engineering
Status
Context
Internal Problem
Staff handling accreditation needed to cross-reference long documents, past applications, and best practice guides simultaneously. It was a slow, manual, error-prone process with a real deadline.
Public Problem
Visitors and members on the Chamber website had no fast way to get answers about events, membership, or services. Missed engagement and repeated staff inquiries were the result.
The Goal
Build two AI systems that solve real organizational problems. Not just demos, but tools people actually use.
What I Built
Internal Tool
An internal AI assistant for Chamber staff over accreditation documents, using a RAG approach with gpt-4o-mini, strict grounding, and fallback handling to prevent hallucinations.
Public-Facing
A FastAPI-powered public-facing website chatbot for visitors and members. Retrieval-grounded with gpt-4o-mini, strict grounding, and fallback handling to prevent hallucinations.
System One · Deep Dive
Input
2026 accreditation question + 2015 completed responses + U.S. Chamber best practices
Processing
Cross-referencing, gap detection, and recommendation generation via engineered AI instructions
Output
Structured 3-part guidance: past answer → current question → recommended approach
Gathered the 2015 completed accreditation application and integrated all supporting attachments: governance, HR, finance, and operational documents.
Manually extracted and restructured accreditation questions from a locked online portal that couldn't be scraped, turning inaccessible content into a clean, usable document.
Added U.S. Chamber best practice example documents, manually merging large PDFs when free online tools failed to handle the file size.
Built detailed prompts guiding the AI to compare questions, detect new requirements, identify gaps in past responses, and suggest stronger, structured answers.
Packaged and delivered the assistant with clear response structure: 2015 MSCC reference → current 2026 question → best practice recommendation.
System Two · Deep Dive
Request Flow
What I Built
FastAPI backend: Request routing, validation, and OpenAI integration
OpenAI API connection: gpt-4o-mini responses scoped to official MSCC content
Structured knowledge files: Newsletter PDFs and website content processed into a searchable knowledge base
Conversation memory: Short-term context window for multi-turn follow-ups
Chatbot bubble UI: Floating React interface designed for website embedding
Scope enforcement: System-level instructions prevent out-of-domain answers
Chatbot Capabilities
Upcoming events, registration details, dates, and sign-up info.
Benefits, tiers, how to join, and what membership includes.
Programs, services, and tools available to Chamber members.
Office address, contact info, and how to reach the Chamber.
What MSCC is, its mission, and how it serves the community.
Out-of-scope queries handled cleanly with no hallucinations.
Engineering Reality
Converting scattered website content, PDFs, and organizational documents into a clean, queryable structure required careful curation and judgment about what to include.
The 2026 accreditation questions lived behind a locked portal that couldn't be scraped. Manual extraction and restructuring was the only viable path.
Free tools couldn't handle the volume of large best-practice PDFs. Manual merging, reformatting, and document organization added significant prep work.
Getting the bot to produce consistently structured and useful output required many iterations of prompt engineering, balancing depth with precision.
Preventing the chatbot from answering out-of-domain questions without breaking conversational flow required careful system instruction design and testing.
Debugging conversation memory between the React frontend and FastAPI backend: session state, context passing, and edge cases in multi-turn dialogue.
Tools & Technologies
Results
AI Systems
Two AI assistants built for real organizational use, not demos.
Manual Work
Reduced manual document lookup for accreditation staff significantly.
Shipped
Public chatbot reached a deployment-ready stage for visitors and members.
Staff Testing
Internal tool delivered and tested against a real accreditation deadline.
This project applied AI to a live organizational workflow, not a classroom exercise. The systems were designed with real users, real data, and real deadlines as constraints. That distinction matters.
Reflection
There's a large gap between "this could work with AI" and building something people trust and use. System design, UX, and edge case handling matter just as much as model choice.
Restricting what a chatbot answers is not a weakness. It's what makes it reliable in production. A well-scoped assistant consistently outperforms a general one in organizational contexts.
No amount of prompt engineering compensates for poorly organized knowledge. The work of curating, structuring, and cleaning content is foundational, not optional.
Want to see more?
I'm open to AI engineering roles, research collaborations, and real-world AI systems projects.