Unsloth AI, GRPO, HF, transformers, PyTorch, LLM optimization, MLOps, OpenCV, GazeNet, WebRTC, FastAPI, PostgreSQL
Worked with the founder Siddhant Bansal and Dr. Tanmoy Chakraborty (both from IIT Delhi) to design proprietary algorithms and LLM optimization techniques. AI interviews for ~100 candidates simultaneously running on local servers.

• Curated the dataset from 5k+ real interview transcripts and synthetic reasoning.
• Fine tuned and applied GRPO on Qwen2.5 3B (4bit) using Unsloth AI.
• Batched requests during post interview analysis for faster inference.
• Async parallelism using asyncio, achieved response time of ~1.5 seconds.
• Debugging stack with system and session level logging.

Product Overview

Griphic Team

PyTorch, YOLO, Dlib, OpenCV, C++, DNN, Python
Running on little RPi4, this AI model can detect if a driver is sleepy, distracted, drowsy, or drinking while driving. Intended to be used by fleet owners to monitor their drivers and prevent disasters. Analytics can be seen in real-time on dashboard.

• Used Dlib 5-point face landmark predictor for fast inference, 54 FPS on RPi 4.
• Custom trained a binary classifier to predict closed vs open eyes.
• SolvePnP algorithm to detect if a driver is not looking at the road.

Demonstration of DMS

Our DMS at Convergence, Delhi

deepstream, C++, OpenCV, GStreamer, Pytorch, CUDA, Python, JavaScript
Powered by Jetson Nano, PixHawk, and 5G module, these drones can be assigned missions, run object detection models, and can be monitored live.

• Trained and optimized object detection pipeline at 80+ FPS using deepstream.
• Engineered the entire system and GCS to stream live camera feed in HD using GStreamer, AI analytics, and telemetry data with < 200ms delay.
• Communicated with drones via telemetry and SSH tunneling.

Our drone flying over Ajmer Road

Ground Control Station (GCS)

OpenCV, C++
Drone is equipped with a sensor that detects gas leakage (green laser inside circle). This laser pointer has to be on the pipeline to work accurately, but it fluctuates due to drone's unstability. The non-AI algorithm uses Lucas-Kanade, Match Template, and noise reduction algorithms to detect and track pipelines. Then it calculates the relative angle (red number in video) that needs to be given to the gimbal to put laser pointer back on the pipeline. Intended to assist drone pilots so that they don't have to correct gimbal frequently.

Demo of Gas pipeline tracker

Detection algorithm

OpenCV, Python, Numpy
The main frame demonstrates the position of drone, with grids representing 1x1 meter in real world. Top right frame is simulating the tracked car's position, bottom right is drone's altitude.
The algorithm then calculates the relative angle and relative speed that needs to be given to the drone to follow the car.

Demonstration of Tracking algorithm

Selective car tracking

OpenAI, MySQL, PHP, JS, EC2, Route53, Cloudfront
Launched this website way before ChatGPT and all the Gen AI hype. Fine tuned text-davinci-003 on instructions dataset and used few shot examples to generate code based on the given prompt.

Ranks #1 for "Exact Codes"

1600+ Monthly Active Users

Tensorflow, OpenCV, Python
Trained an AI model from scratch to detect asteroids in telescopic images (.fits) distributed by IASC.
The model detected three Asteroids out of which two were officially given preliminary status by NASA.

Explaination on my channel

Certificate