I deleted Instagram more than a year ago, and honestly, it saved me from a lot of distractions.
But there was an unexpected downside.
A lot of informal, real-time information β especially during college events β still lives there.
During our college fest, for example:
- Event schedules
- Last-minute updates
- Food stall announcements
- Informal activities
β¦everything gets posted on Instagram.
At the same time:
- Fest details and significance are on the official website
- Food stall info is on a separate app
- The entire 3-day schedule is compressed into a few posts
Thereβs no single place to get a clear, structured view of everything.
And thatβs when it hit me:
Most college fests have websites.
Some even have apps.
But none of them actually help you navigate the fest intelligently.
They give information.
They donβt give guidance.
But I wanted to build something smarter β
an AI assistant that actually understands queries, plans your day, and even helps you find teammates.
So, I built Pragyan Mentor Assistant β an AI-powered system for navigating a techno-managerial fest.
π― Problem
During college fests like Pragyan (NIT Trichy):
There are
- There are dozens of events, workshops, and shows
- Information is scattered across PDFs, sites, and posters
-
Users donβt know:
- what to attend
- what matches their interests
- how to plan their time
- who to team up with
π Traditional apps = static information
π I wanted intelligent interaction
π‘ Solution
I built a multi-tool AI assistant that can:
- π Answer questions about events, workshops, proshows
- π Show food stalls & mess menu
- π§ Recommend activities based on user intent
- π Plan your schedule
- π€ Match you with like-minded participants/Suggest potential teammates (prototype)
- π Answer fest-related questions using RAG
π§ System Design
Instead of a simple chatbot, I designed it as a tool-using agent system.
πΉ Tools
fetch_eventsfetch_workshopsfetch_food_stallfetch_mess_menu-
pragyan_bot(RAG-based) smart_recommenderplannerbuddy_matcher
πΉ Agent Flow
- User query
- LLM decides:
- Which tool to call
- Tool executes
- Response is generated in natural language
π Retrieval Approach
This system uses a hybrid retrieval strategy at the system level:
-
Structured retrieval (keyword-based)
- Direct tool calls for events/workshops
- Fast and deterministic
-
Semantic retrieval (RAG)
- Vector search over fest documents
- Handles open-ended queries
π This combination allows both precision and flexibility
π RAG (Retrieval Augmented Generation)
To handle fest knowledge:
-
Used:
- Text files (events, shows, lectures, FAQs)
-
Built:
- FAISS vector store
-
Retrieval:
- Semantic search on query
-
Response:
- Context-aware answers
π§ Memory
Using:
-
InMemorySaver()(LangGraph)
π Enables:
- remembering user preferences
- better recommendations
- conversational continuity
π€ Smart Features
π― Recommendations
Understands intent like:
"What should I attend if I like tech and fun?"
π Planner Agent
"Plan my next 3 hours"
Generates a structured schedule based on:
- time
- interests
- available events
π€ Buddy Matching (Prototype)
Matches based on:
- interests
- level
- context (e.g. case study competitions)
Uses a small dataset to demonstrate logic
π₯οΈ UI
Built with Streamlit:
- Chat-based interface
- Quick action buttons
- Structured responses
π Deployment
Deployed on Render (free tier)
Environment variables for API security
π₯ Demo
π https://www.loom.com/share/13f87025a9154a55b80fc240bfc91ba2
π οΈ Tech Stack
- Python
- LangChain
- OpenAI API
- FAISS
- Streamlit
- Render
β οΈ Challenges Faced
- RAG retrieval quality (chunking + parsing issues)
- Tool selection accuracy
- Structuring multi-agent workflow
- Deployment + API key handling
π Ongoing Improvements
Some features Iβm actively working on:
- Adding database-backed user profiles for real buddy matching
- Improving RAG with better retrieval and evaluation
- Expanding dataset coverage for more complete fest information
- Exploring true hybrid retrieval + reranking
π What I Learned
- Building agents > building chatbots
- RAG needs data structuring, not just embeddings
- UI matters a lot for perceived intelligence
- Deployment and debugging are part of the real challenge
π Links
- π Live demo available on request
π Final Thoughts
This project made me realize:
π The future isnβt just about LLMs
π Itβs about systems built around them
If you have suggestions or ideas to improve this, Iβd love to hear them!









![Defluffer - reduce token usage π by 45% using this one simple trick! [Earthday challenge]](https://media2.dev.to/dynamic/image/width=1000,height=420,fit=cover,gravity=auto,format=auto/https%3A%2F%2Fdev-to-uploads.s3.amazonaws.com%2Fuploads%2Farticles%2Fiekbgepcutl4jse0sfs0.png)


