DEV Community: VICTOR LAKRA

AI-Powered Teacher Assistant: Revolutionizing Lesson Planning for Educators

VICTOR LAKRA — Sun, 28 Sep 2025 22:56:01 +0000

This is a submission for the Heroku "Back to School" AI Challenge

What I Built

I built Teacher Assistant, a web application that helps educators generate structured lesson plans in minutes. Teachers often spend significant time every day designing lesson plans, which can be repetitive and time-consuming. This application leverages AI to streamline the process: educators simply upload the document they intend to teach (e.g., a textbook chapter, reading material, or notes), and the system generates a context-aware lesson plan tailored to the grade level, topic, and duration specified by the teacher.

The generated lesson plan includes:

Step-by-step teaching activities
Required materials
Timed sections for better pacing
Formative quiz questions with answers
Differentiation for two levels (support & extension)

This helps teachers save time and focus more on meaningful student interaction rather than paperwork.

Demo

Checkout the project here: Teacher Assistant

Flow of the application:

Teacher uploads a PDF document of the lesson content.
The system extracts text and creates vector embeddings, stored in a pgvector-powered database.
Teacher specifies topic, grade and duration.
With one click, a tailored lesson plan is generated, aligned with the uploaded material.

Screenshots

Screenshots showcasing the Teacher Assistant UI, document upload workflow, and generated lesson plans.

How I Used Heroku AI

The application integrates multiple Heroku AI features:

1. Heroku Managed Inference and Agents

Embeddings Generation: Used Cohere embeddings API to convert uploaded document text into embeddings.
Lesson Plan Generation: Used Claude Sonnet 4 to generate structured lesson plans based on both teacher input (grade, topic, duration) and retrieved contextual embeddings.

2. pgvector for Heroku Postgres

All embeddings are stored and queried in a pgvector-enabled Postgres database.
This ensures semantic search capabilities to retrieve the most relevant content from the uploaded material during lesson plan generation.

Together, these components enable a seamless multi-agent workflow: one agent handles document embedding and storage, while another focuses on contextual lesson plan generation.

Technical Implementation

Frontend: React + TypeScript, with Tailwind CSS for styling.
Backend: Express + TypeScript
AI Workflow:
- On document upload, text is extracted and embeddings are generated via Managed Inference (Cohere).
- Embeddings are stored in Postgres with pgvector extension for efficient semantic retrieval.
- When the teacher requests a lesson plan, embeddings are queried to extract the most relevant context, and an agent (Claude Sonnet 4) generates the structured plan.
Multi-Agent Architecture:
- Document Agent: Handles embedding creation and storage.
- Planning Agent: Generates the lesson plan using retrieved context and teacher inputs.
Database: Heroku Postgres with pgvector.

Challenges Solved

Efficient context retrieval: Implemented semantic search over lesson documents using pgvector to ensure lesson plans are accurate and aligned with actual teaching material.
Multi-agent coordination: Designed a workflow where one agent enriches the knowledge base (embeddings) and another produces actionable lesson plans.
User experience: Created a minimal, intuitive interface where teachers can generate professional-grade lesson plans with a single click.

Results

I have pitched and demoed Teacher Assistant to few educators. The response was overwhelmingly positive, several teachers expressed strong interest in adopting it immediately, finding it a valuable tool to save time and improve teaching quality. The application demonstrates how AI can meaningfully empower educators while improving classroom experiences.

From Iron Man to n8n: Creating a JARVIS-Like AI Assistant

VICTOR LAKRA — Mon, 01 Sep 2025 02:29:31 +0000

This is a submission for the AI Agents Challenge powered by n8n and Bright Data

What I Built

I built a JARVIS-inspired AI Agent workflow in n8n that acts as a smart assistant capable of handling two core tasks:

Fetching the latest news using a dedicated news_agent workflow.
Analyzing Crunchbase startups/companies based on keyword-driven queries via the analyze_startups workflow.

It also includes a fun wake-up activation feature: when the user says "Wake up, daddy's home." it triggers music playback from Spotify, just like summoning JARVIS in Iron Man style. This combines AI-powered intelligence with practical and entertaining automation.

Demo

Here’s a walkthrough of the JARVIS AI Agent in action:

You can issue a query like “Analyze fintech companies on Crunchbase” and the workflow extracts the keyword and calls the analyze_startups sub-workflow.
A query like “Get me the latest news” routes to the news_agent workflow, which scrapes BBC, enriches with Bright Data, and formats results into a witty newsletter-style digest.
Saying “Wake up, daddy's home.” makes it greet you by starting a Spotify playlist.

n8n Workflow

You can view the full workflow JSONs here:

Jarvis Main Workflow

News Agent Workflow

Analyze Startups Workflow

Technical Implementation

The agent is built using n8n’s AI and automation nodes:

System Instructions: A strict prompt ensures JARVIS only handles news and Crunchbase analysis, rejecting unrelated queries.
Model Choice: gpt-4o-mini powers the conversational reasoning.
Memory: A buffer memory node maintains conversational context per session.
Tools: Two tool workflows — news_agent for news, analyze_startups for startup analysis.
Webhook: Handles user input and responds back with AI-driven results.
Spotify Node: Adds the interactive wake-up feature for user delight.

News Agent

Pulls BBC headlines via RSS.
Uses Bright Data to scrape and enrich article content.
Aggregates and formats results.
Sends to an AI Agent node with instructions to rewrite in a Morning Brew–style witty newsletter.
Delivers the final email digest via Gmail integration.

Analyze Startups Agent

Triggers a Bright Data Crunchbase dataset search based on a keyword (e.g., fintech, AI, climate tech).
Polls until the snapshot is ready.
Retrieves startup/company details.
Cleans and sorts data with Python code (top 10 by most recent founded date).
Returns a concise, structured summary with insights on funding, team, products, and market signals.

Bright Data Verified Node

Both news_agent and analyze_startups integrate Bright Data’s verified data nodes for reliable scraping and enrichment. This ensures that the insights pulled from external data sources are accurate, fresh, and robust.

Why Real-Time Data Mattered

Real-time data made JARVIS genuinely useful instead of just being a static demo. Fresh headlines gave the news digest actual relevance for the day, and pulling Crunchbase startups dynamically ensured users could explore current market players rather than outdated datasets. Without real-time enrichment, the workflows would have felt like canned responses instead of a living, adaptive assistant.

Journey

The journey started with wanting to create a JARVIS-like AI that’s both useful and fun. I faced challenges in:

Designing a clear system prompt so the AI never strays outside its defined scope.
Ensuring keyword extraction for Crunchbase queries worked consistently.
Managing session memory to give the agent contextual awareness.
Formatting news into a polished HTML digest that feels engaging.

Key learnings:

n8n’s agent node makes it seamless to route user queries to the right tools.
Bright Data’s verified node is essential for scaling reliable web data into AI pipelines.
Adding small touches like Spotify integration creates a delightful user experience.

This project brought together AI, automation, and a touch of sci-fi imagination into one cohesive workflow.

DEV Community: VICTOR LAKRA

AI-Powered Teacher Assistant: Revolutionizing Lesson Planning for Educators

What I Built

Category

Demo

Screenshots

How I Used Heroku AI

Technical Implementation

Challenges Solved

Results

From Iron Man to n8n: Creating a JARVIS-Like AI Assistant

What I Built

Demo

n8n Workflow

Technical Implementation

News Agent

Analyze Startups Agent

Bright Data Verified Node

Why Real-Time Data Mattered

Journey