kashish saini

AI/ML Developer

Gurugram, Haryana, India 0+ yrs exp 80 · Excellent

About

B.Tech CSE graduate (CGPA 8.10/10) and AI/ML Intern building and benchmarking deep generative models (VAE, GAN, Diffusion) in PyTorch for synthetic data generation. Evaluated four VAE variants, DCGAN and DDPM on MNIST, CIFAR10, retail and chest X-ray images using FID, LPIPS and Inception Score, and measured downstream gains with YOLOv8 classifiers. Also built two full-stack LLM applications with FastAPI, React and Groq (Llama 3.3). Seeking an entry-level AI/ML Engineer or Generative AI Developer role.

Skills & Expertise (21)

Python Advanced

9.0/10

Years Exp

PyTorch Advanced

9.0/10

Years Exp

Prompt Engineering Advanced

8.0/10

Years Exp

Kaggle Advanced

8.0/10

Years Exp

Object-Oriented Programming Intermediate

7.0/10

Years Exp

Data Structures and Algorithms Intermediate

7.0/10

Years Exp

Git Intermediate

7.0/10

Years Exp

SQLAlchemy Intermediate

7.0/10

Years Exp

MongoDb Intermediate

7.0/10

Years Exp

SQLite Intermediate

7.0/10

Years Exp

MySql Intermediate

7.0/10

Years Exp

CSS3 Intermediate

7.0/10

Years Exp

HTML5 Intermediate

7.0/10

Years Exp

Redux Toolkit Intermediate

7.0/10

Years Exp

React.js Intermediate

7.0/10

Years Exp

REST APIs Intermediate

7.0/10

Years Exp

Django Intermediate

7.0/10

Years Exp

FastAPI Intermediate

7.0/10

Years Exp

C++ Intermediate

6.0/10

JavaScript Intermediate

6.0/10

Java Intermediate

6.0/10

Work Experience

AI/ML Intern, Synthetic Data Generation

Truminds Software Systems Pvt. Ltd.

Aug 2026 - Present

Built and benchmarked 4 VAE variants (Vanilla, CVAE, Beta-VAE, VQ-VAE) in PyTorch on retail and chest X-ray images; tuned learning rate, latent size, beta and codebook parameters with Optuna, raising downstream YOLOv8 accuracy from 16.2% to 17.0%. Trained DCGAN and DDPM (U-Net, EMA, DDIM sampling) on MNIST and CIFAR10; stabilised GAN training (label smoothing, gradient clipping) after fixing a discriminator/generator collapse. Showed diffusion gave better image quality: mean FID 88.2 (DDPM) vs 137.6 (DCGAN); on MNIST FID 13.2 vs 54-69; scarce-data DDPM reached CIFAR10 FID 110.2 using only 60 real images/class. Quantified data scarcity: training on 60 images/class raised VAE FID from 211.3 to 320.4 (+51.6%). Designed a synthetic-augmentation pipeline: pseudo-labelled generated images with YOLOv8n-cls (99.0% MNIST, 92.2% CIFAR10), compared LPIPS-based vs random selection (60 vs 180 images/class); LPIPS selection cut perceptual distance by 36-67%, and 180 random images gave the best accuracy (84.3-84.5%). Optimised runs for a Kaggle T4 GPU budget with fp16 mixed precision, checkpoint scoring and seeded, reproducible comparisons documented in cross-notebook reports.