ohmymbbs

• 

  elasticsearch agent in action inside elastic dashboard

• 

  telegram agentic bot ui in action https://t.me/Neetpgprepbot

• 

  v1 - previous UI

• 

  v2 agentic UI with tool calling

• 

  v2 and instant feedback

• 

  telegram action n elastic search tool call based on action button chosen

• 

  API for future swiftui widget/other integrations. JSON render to see error, test and build other tool based UI workflows

• 

  agent working inside claude reusing most of the current code and MCP

• 

  claude-settings-connector add MCP url to see in action https://neetpg-chatgpt-app.vercel.app/mcp

Inspiration

I built this for my sister, an MBBS graduate preparing for NEET-PG. Her biggest pain wasn’t content shortage, it was friction: opening multiple apps, losing context, and no fast way to practice in micro-sessions. its nowhere near Marrow but a definite start.

What it does

• Delivers adaptive MCQ practice by subject/topic/random.
• Supports chat + interactive UI cards (not just plain text).
• Tracks attempts, accuracy, latency, and weak topics.
• Works across channels: web and Telegram-style quick actions.
• Shows progress summaries so users can continue from where they left off.

How we built it

• med-mcqa-data for question bank. (library containing 180k Q/A database)
• Built with Elastic Agent Builder + custom tools including keyword search, filtered MCQ, topic and distribution analytics
• Frontend uses JSON-rendered adaptive blocks: mcq_card, answer_feedback, progress_stats, weak_topics, quick_actions

Accomplishments that we're proud of

text clunky. and no difference as to chatgpt. like everytime i went for feedback, she was like chatgpt can do that. with UI and agents i have given her hope and something to try. requires lots of work and ofcourse way to figure out production, 14 days limit and more. but it was frictionless with codex and claude code at helm with mcp.

What's next for ohmymbbs

• Difficulty calibration using model-based benchmarks:
• compare how older/newer models solve each question
• rank questions by cognitive difficulty and misconception risk when critical mass uses our app.
• Personalized daily plans (time-boxed revision + weak-topic reinforcement).
• Better UI layers: richer progress journeys, streaks, and concept maps.
• More channels and integrations while preserving one analytics backbone.
• Closed-loop recommendations: from “what you got wrong” to “what to do next today.”

Built With

• agentic-ui
• elasticsearch
• json-render
• tailwind
• telegram
• vercel