Harshal Poriwade

AI/ML Engineer specializing in Computer Vision, Deep Learning, and production-grade systems.

Building intelligent systems with Python, PyTorch, TensorFlow, and FastAPI. Currently exploring Large Language Models and transformer architectures.

About Me

Computer Science graduate with proven expertise in Machine Learning, Deep Learning, and Computer Vision. Experienced in building and deploying production-grade AI systems using PyTorch, TensorFlow, and FastAPI. Co-authored published research on multi-modal deepfake detection achieving AUC 0.9958. Passionate about solving real-world problems through data-driven solutions, scalable API development, and edge deployment. Currently exploring Large Language Models and transformer-based architectures.

Experience

AI/ML Intern
SCRYMZ Private Software Pvt. Ltd. • March 2026 – Present
  • ▸Architected a production-grade multi-model Computer Vision pipeline using YOLOv8, Vision Transformers (ViT), and custom CNNs for automated structured data extraction from complex, real-world images.
  • ▸Trained a custom CNN classifier on heavily stylized graphical elements achieving 99.91% accuracy, entirely bypassing traditional OCR limitations.
  • ▸Custom-trained a deep learning Vision Transformer-based OCR model on 30,000+ synthetic images achieving 90.16% sequence accuracy and 97.23% character accuracy.
  • ▸Engineered a FastAPI backend deployed via Gunicorn/Uvicorn exposing 5 modular REST endpoints for asynchronous batch processing validated with Pydantic.
  • ▸Optimized a 5-model AI pipeline for CPU-only AWS deployment achieving sub-2-second end-to-end latency via tensor batch parallelization.
Python
PyTorch
TensorFlow
FastAPI
YOLOv8
Vision Transformers
AWS
Gunicorn/Uvicorn

Projects

DeepFake Detection System
Architected a real-time multi-modal deepfake detection system supporting Video, Image, Audio, and Text inputs with a FastAPI backend exposing 8 RESTful endpoints. Engineered EfficientNet-B4/B5 ensemble for video, custom audio CNN with MFCC, and HuggingFace Transformers for NLP fact-checking. Achieved AUC 0.9958 and F1-score 0.9726 with <40ms inference; co-authored published research in Journal of Cyber Security.
Python
PyTorch
TensorFlow
FastAPI
EfficientNet
HuggingFace Transformers
Spectrogram-UNet-Denoiser
Engineered a 4.48M parameter Frequency-Domain 2D Spectrogram U-Net matching denoising quality of 15M+ parameter models, processing 48kHz audio via STFT with Log1p compression. Built dynamic tf.data pipeline with 16,000 unique audio slices per epoch and two-stage training on extreme SNR [-15dB to 10dB]. Achieved +5.41dB average SNR improvement and deployed as 17.1MB TFLite model with <10ms inference via XNNPACK.
TensorFlow
TFLite
Python
Audio DSP
Librosa
STFT
Car Price Predictor
Designed and implemented a machine learning model to predict the resale price of used cars based on key features such as brand, model, year, and fuel type. Used a cleaned dataset to train regression models including Linear Regression, optimizing for accuracy and generalization. Integrated the model into an interactive web app using Streamlit for real-time predictions.
Python
Scikit-learn
Pandas
Streamlit
Jupyter Notebook
Movie Recommendation System
Developed a machine learning-based recommendation engine for generating personalized movie suggestions. Utilized the TMDB 5000 Movies and Credits datasets for constructing a comprehensive content and metadata-based recommendation pipeline. Implemented multiple feature extraction and preprocessing steps to capture relevant movie attributes.
Python
Scikit-learn
Pandas
Jupyter Notebook
Audio Translation System
Designed an audio-to-audio translation tool using speech recognition, text-to-speech, and translation APIs. Supports live input, file upload, and language selection for seamless multilingual communication.
Python
Speech Recognition
TTS APIs
Translation APIs

Skills

Languages

Python

Deep Learning & Computer Vision

PyTorch
TensorFlow
Keras
TFLite
Vision Transformers (ViT)
YOLOv8
EfficientNet
U-Net
Grad-CAM (XAI)

Machine Learning

Scikit-learn
SHAP
Ensemble Learning
Feature Engineering
Hyperparameter Tuning

Frameworks & Libraries

FastAPI
Gunicorn
Uvicorn
OpenCV
Librosa
HuggingFace Transformers
Pandas
NumPy

Deployment & Optimization

TFLite Quantization
AWS
Edge Deployment
Model Optimization
Pydantic
Swagger UI

Tools & Platforms

Git
GitHub
Roboflow
Google Colab
Kaggle
Jupyter Notebook
Postman

Databases

MySQL

Education

B.E. in Computer Science and Engineering

Angadi Institute of Technology and Management, Belgaum, Karnataka

Dec 2022 – July 2026

CGPA: 7.8

Achievements

1st Place – Tech Project Showcase
SGBIT

March 2025

2nd Place – IEEE Project Exhibition, INGENIOUS
AITM

April 2025

3rd Place – 24-Hour Hackathon, TCS Git
GIT

June 2025

Co-Author – Multi-Modal DeepFake Detection using Ensemble Learning
Published in Journal of Cyber Security

2025