DEV Community: Naitik Verma

Help me learn Ai agents.

Naitik Verma — Sat, 07 Mar 2026 03:46:28 +0000

This is not exactly a post Just want to ask I am starting my journey learning AI agents what would be the best roadmap or resources from where i can learn and do practical implementation for free.

Built a RAG AI Teaching Assistant for Video Lectures

Naitik Verma — Sat, 07 Mar 2026 03:28:12 +0000

Most of the time when watching long lecture videos, finding a specific concept later becomes difficult. You either have to rewatch the entire lecture or manually search through timestamps. I wanted a system where you could simply ask questions about a lecture video and get answers instantly.

So I built a Retrieval-Augmented Generation (RAG) based AI Teaching Assistant for video lectures.

The idea is simple: convert lecture videos into searchable knowledge.

Pipeline:

Video → Audio
The lecture video (MP4) is first converted into audio (MP3).

Audio → Transcript
The audio is transcribed into text so the system can understand the lecture content.

Chunking + Embeddings
The transcript is split into smaller chunks and converted into embeddings.

Vector Retrieval
The embeddings are stored in a vector index. When a question is asked, the system retrieves the most relevant lecture segments.

LLM Answer Generation
The retrieved lecture context is passed to a local LLM running with Ollama, which generates the final answer grounded in the lecture content.

This has the capability to reduce traffic and pollution with densely populated countries.

Naitik Verma — Fri, 06 Mar 2026 19:05:15 +0000

AI-Based Green Light Optimization using Computer Vision

Naitik Verma ・ Mar 6

#ai #deeplearning #machinelearning #showdev

AI-Based Green Light Optimization using Computer Vision

Naitik Verma — Fri, 06 Mar 2026 18:58:26 +0000

Urban traffic systems still rely largely on fixed timer traffic lights. These timers do not adapt to real-time traffic conditions, which often leads to congestion, unnecessary waiting time, and increased fuel consumption.

To explore a more intelligent approach, I built the Metropolitan AI Control Center, a traffic signal optimization system that combines Deep Reinforcement Learning, Computer Vision, and traffic simulation.

The goal of the project is to replace static traffic signal timers with an AI agent that continuously learns how to manage intersections based on real-time traffic conditions.

Project Overview

The system operates on a simulated 10-intersection city grid using Eclipse SUMO (Simulation of Urban Mobility). A Deep Q-Network (DQN) agent learns how to control signal phases in order to reduce overall waiting time and prevent congestion from spreading across intersections.

Vehicle density is estimated using YOLOv8-based computer vision, while the entire system is monitored through a Flask-based web dashboard.

This setup allows the AI to interact with a realistic traffic environment and continuously improve its signal control strategy.

Technology Stack

AI and Learning

Python 3.9+

PyTorch

Deep Q-Network (DQN)

Computer Vision

Ultralytics YOLOv8 for vehicle detection and classification

Traffic Simulation

Eclipse SUMO

TraCI API

Web Interface

Flask

HTML5

CSS3

JavaScript

Analytics

NumPy

Matplotlib

Key Features
Deep Reinforcement Learning Signal Control

A Deep Q-Network agent learns optimal traffic signal policies by interacting with the SUMO simulation environment. The objective is to minimize waiting time across all intersections while preventing traffic spillback.

Computer Vision Traffic Monitoring

Simulated camera feeds are processed using YOLOv8, which detects different types of road users such as cars, trucks, buses, bikes, and pedestrians. This information is used to estimate traffic load.

Emergency Vehicle Priority

The system includes logic to detect emergency vehicles such as ambulances and temporarily prioritize their routes by adjusting signal phases.

Environmental Impact Monitoring

The simulation also tracks metrics such as:

CO2 emissions

fuel consumption

acoustic noise levels

This helps evaluate the environmental impact of improved traffic flow.

Comparative Testing Mode

The dashboard includes a testing feature that compares traditional traffic control with the AI system.

Phase 1 – Baseline Simulation
A SUMO simulation runs with traditional fixed traffic timers.

Phase 2 – AI Optimization
A new isolated SUMO instance runs with the trained DQN agent controlling the signals.

The dashboard then visualizes differences in:

average waiting time

traffic flow efficiency

CO2 emissions

Repository

https://github.com/naitik23verma/Green_light_optimization