About Me
Rochester, NY | +1 (585) 576-1904 | parasjain1999@gmail.com | LinkedIn | GitHub
Education
-
M.S. in Artificial Intelligence, Rochester Institute of Technology
GPA: 3.9/4.0, Aug 2023 - May 2025
Courses: High-Performance Architecture, Robotics Software Systems, Computer Vision, Machine Learning, CUDA
-
B. Tech in Computer Science, SRM Institute of Science and Technology
GPA: 8.5/10.0, July 2017 - May 2021
Skills and Technologies
- Programming Languages: Python, C++, C, Bash, R, SQL
- ML/AI Frameworks: PyTorch, TensorFlow, Hugging Face, Scikit-learn, Lang Chain, OpenCV, CUDA
- Cloud & Tools: Azure ML, AWS, Fast API, Git, Jenkins, FAISS, Tru Lens, Deep Eval, MongoDB, VLM
- Core Expertise: Machine Learning, Deep Learning, Computer Vision, LLM Evaluation, NLP, Reinforcement Learning
Work Experience
-
AI Intern, NetApp
May 2024 – Dec 2024
- Built a RAG-based chatbot system integrating FAISS vector search with OpenAI LLMs, achieving accurate document-grounded Q&A with zero hallucination on internal knowledge base
- Deployed Azure ML-powered Auto ML pipelines for ARR forecasting, achieving over 95% accuracy on quarterly revenue predictions using time series analysis and feature engineering
-
Software Developer, Barclays
Aug 2021 – Aug 2023
- Built ETL screening engine for Anti-Money Laundering (AML) detection, improving accuracy by 25% through data quality optimization and feature engineering
Academic Projects
-
High-Resolution Segmentation via SAM and Super-Resolution
Sep 2024 – Dec 2024
- Augmented Meta’s Segment Anything Model (SAM) with a Laplacian pyramid-based super-resolution module to enhance spatial fidelity and improve semantic segmentation performance under degraded visual inputs.
- Integrated a multi-stage pipeline where high-frequency reconstruction preceded segmentation, enabling sharper attention map focus across the encoder-decoder transformer backbone.
- Quantitatively benchmarked using IoU, Dice, and pixel-wise F1 metrics, showing up to +12% gain in boundary-sensitive segmentation tasks.
-
Atrial Fibrillation Detection Using Transformer Models on ECG Signals
Dec 2020 – May 2021
- Built a deep learning pipeline using Transformer encoders to classify Atrial Fibrillation episodes directly from long-duration 1D ECG signals, leveraging temporal attention to capture irregular heartbeat patterns.
- Preprocessed biomedical time-series data into normalized and windowed signal segments using sliding windows aligned with AFib annotations.
- Trained models on multiple patients’ ECG recordings using label-balanced sampling and positional encoding, achieving high classification performance with low false positives.
- Demonstrates expertise in sequence modeling, time-series, transformer architectures, and biomedical ML.
-
CUDA Kernel Optimization for VGG16
Jan 2025 – May 2025
- Architected custom CUDA C++ kernels implementing shared memory tiling, coalesced memory access, and thread block-level parallelism for 2D convolutional layers from VGG16.
- Achieved up to 7× speedup over naive GPU implementations, validated against PyTorch ONNX inference outputs for L2-norm error stability.
- Analyzed memory throughput using CUDA profiler (Nsight) and optimized kernel occupancy through loop unrolling and bank conflict minimization.
- Demonstrates deep competency in model-level hardware acceleration, GPU-bound CV workloads, and on-device neural inference optimization, critical for edge-based XR rendering.
-
Indic OCR Enhancement Using Vision-Language Model Fine-Tuning
Feb 2024 – Apr 2024
- Fine-tuned Qwen2-VL, a multimodal vision-language transformer, for low-resource Gujarati OCR, using a custom dataset comprising scanned documents and native script annotations.
- Preprocessed image inputs with OpenCV morphological operations and Real-ESRGAN for resolution recovery, improving input clarity under noisy scan conditions.
- Evaluated with WER (Word Error Rate) and CER (Character Error Rate), achieving significant reductions post-enhancement.
- Demonstrates capability in ground truth dataset generation, multilingual vision-language fine-tuning, OCR pipeline curation, and low-resource domain adaptation directly aligned with spatial AI localization and meta-linguistic modeling.
-
Facial Emotion Recognition via Landmark-Guided CNNs (Published)
Dec 2020 – May 2021
- Developed a facial effect classification pipeline using statistically normalized landmark-based ROI extraction and a compact CNN classifier optimized for real-time inference.
- Published results in an IEEE venue, emphasizing generalizability and label alignment robustness across cultural expression variances.
Publications
-
Face Emotion Detection Using Deep Learning
2021 Fifth International Conference on I-SMAC (IoT in Social, Mobile, Analytics and Cloud) (I-SMAC), Palladam, India, 2021, pp. 517-522, doi: 10.1109/I-SMAC52330.2021.9641053
-
Human Machine Confidence Comparison
IEEE CogMI (Under Review, June 2025)