Translational Natural Language Inference (NLI)
Built a translation-based NLI pipeline using transformers, achieving strong multilingual inference results.
- Hugging Face Transformers
- PyTorch
- Seq2Seq Translation
- Multilingual NLP
- Model Evaluation
Hi — I’m
Guruprasanna Rajukannan Suresh
I build and deploy intelligent AI systems from training models and fine-tuning foundation AI to scalable, production-ready applications.
Scroll
Story About Me
Since childhood, I’ve been fascinated by how the brain turns information into knowledge and action. That curiosity led me to study Electronics, where circuits felt like neurons, and later into automation engineering, building data pipelines, dashboards, and predictive maintenance systems that helped machines anticipate failures.
Wanting to go deeper into intelligence, I pursued my Master of Engineering at the University of Waterloo, focusing on Computational Intelligence, Deep Learning, NLP, Image Processing, and Statistical Modelling. This gave me the foundation to bridge research ideas with production-ready AI.
Since then, I’ve worked across healthcare, accessibility, government, and automation from fracture detection models and speech systems to large-scale LLM deployments for social good. My passion lies in turning research prototypes into reliable, scalable AI systems, blending technical depth with pragmatic engineering. Check my projects or get in touch.
Projects
FashionVision: Multi-Architecture CNN Analysis
Implemented and compared 10 CNN models on Fashion-MNIST, achieving 93.8% accuracy and extracting features for clustering and label discovery.pplicable.
- PyTorch
- CNNs
- PCA, t-SNE, Isomap
- Clustering (DBSCAN, K-means)
2D to 3D Image Reconstruction
Built a 3D reconstruction pipeline from 2D CT scans using feature detection and matching, achieving 90%+ accuracy in medical image analysis.
- OpenCV
- RANSAC Filtering
- 3D Reconstruction (Point Clouds)
- Medical Image Processing
Image Denoising with Convex Optimization
Implemented convex optimization techniques for image denoising, improving restoration quality across multiple noise models.
- Convex Optimization
- Image Denoising
- Noise Models
- PSNR Evaluation
NeuralNet Toolkit: CNN, RNN & NLP Implementations
Built and compared CNN, RNN, and NLP models for image classification, time-series forecasting, and sentiment analysis.
- PyTorch
- Time-Series Forecasting
- Sentiment Analysis
Dimensionality Reduction for Classification Performance
Applied PCA and LDA to improve model efficiency and evaluated trade-offs across multiple classifiers.
- Dimensionality Reduction
- Cross-Validation
- Hyperparameter Tuning
- Gradient Boosting
FasterRCNN Embed Extract
Built an embedding extraction pipeline with Faster R-CNN, enabling custom classifiers, clustering, and similarity search on images.
- Faster R-CNN
- Embedding Extraction
- Object Detection
- NumPy
KSOM: Kohonen Self-Organizing Map
Implemented a Self-Organizing Map for RGB clustering, producing well-structured visualizations of color similarity.
- Self-Organizing Maps
- Clustering
- Unsupervised Learning
- Learning Rate Scheduling
Multilayer Perceptron (MLP) Implementation
Built an MLP from scratch in Python/NumPy, achieving 89% accuracy on multi-class classification tasks.
- Python
- Gradient Descent
- Backpropagation
- Neural Networks (MLP)
Radial Basis Function Neural Network (RBF-NN)
Implemented an RBF neural network with Gaussian kernels, achieving up to 99% accuracy on classification tasks.
- RBF Networks
- Gaussian Kernels
- Clustering (K-Means)
- Stability Analysis
Experience
UN-Habitat - Quality of Life Initiative
AI Software Engineer
Developed an AI toolkit for the UN-Habitat Quality of Life Initiative, enabling knowledge retrieval and decision support by connecting LLMs with structured data and external tools. The system delivers faster insights and recommendations to support policy and decision-making.
- LangChain
- LangGraph
- Multi-Agent Systems
- AWS
- Agentic RAG
- Redis-Celery
- FastAPI
- Docker
- Enabled dynamic case routing by implementing a LangGraph agent with Llama-3.3-70B across law, case-study, and web search nodes
- Delivered <30s query times through a ReAct multi-agent system integrating semantic retrieval and SQL execution.
- Reduced processing 1h → 15m by deploying PGVector + MapReduce ingestion pipelines on AWS ECS with Redis-Celery.
HolisticMindAI
AI Developer
Architected and developed the backend for a therapy support platform, integrating speech diarization, semantic search, and RAG pipelines to deliver scalable, AI-powered summaries.
- WhisperX
- AWS Bedrock
- RAG with LlamaIndex
- FastAPI
- CI/CD
- Docker
- AWS
- Improved speech clarity by implementing WhisperX + PyAnnote diarization, separating speakers with 1.5s accuracy.
- Enabled scalable backend delivery with FastAPI endpoints and AWS Bedrock Mistral-8x7B deployment.
- Boosted document search recall from 68% → 92% with a hybrid RAG pipeline using LlamaIndex + spaCy embeddings.
X-Care
Machine Learning Engineer
Developed a fracture detection pipeline for medical imaging, combining deep learning models, optimized training, and automated deployment supporting faster and more reliable clinical decision-making.
- PyTorch
- EfficientNet
- YOLOv8
- Optuna
- DagsHub
- MLflow
- Achieved 94%+ accuracy by building a two-stage detection pipeline with EfficientNet (classification) and YOLOv8 (localization).
- Reduced training time by 60% through Optuna hyperparameter optimization in PyTorch.
- Automated versioning and deployment of 10+ model runs with MLflow + DagsHub pipelines.
Region of Waterloo (Smart Waterloo Region Innovation Lab)
MLOps Engineer (Co-op)
Built and deployed LLM-powered systems for school policy management, fine-tuning models and engineering automated pipelines delivering scalable, production-ready AI solutions for the public sector.
- LoRA / QLoRA
- RAG
- Hugging Face
- CI/CD
- Containerization
- Docker
- Evaluation
- Enabled lightweight inference by fine-tuning Mistral-7B with LoRA and deploying GGUF models on Hugging Face.
- Improved Q&A accuracy (BLEU 0.85) with a RAG-based policy management system deployed across 5+ school boards.
- Streamlined deployments by engineering a CI/CD pipeline with dynamic ports, auto-versioning, and containerization.
Multiversal Automation Private Limited
Automation Engineer
Engineered real-time data pipelines and predictive maintenance models across manufacturing stations, improving efficiency, extending tool life, and reducing rejects.
- Data Pipelines
- XGBoost
- Predictive Maintenance
- Time-Series Modeling
- Industrial IoT
- Pick & Place Robot
- LabVIEW
- Improved efficiency by building a real-time pipeline across 10 stations with multi-sensor traceability.
- Extended tool life by 25% using XGBoost on torque/load data for predictive maintenance.
- Reduced false rejects by 15% with a time-series anomaly detection model on press/flow/torque curves.
Tube Investments Optoelectronic Solutions
Project Engineer
Designed ETL pipelines and real-time reporting systems for manufacturing data, enabling continuous monitoring, faster insights, and data-driven decision-making.
- ETL Pipeline
- SQL
- Full-Stack Prototyping
- Data Visualization
- Tableau
- Automation
- Coating
- Improved data visibility by designing ETL pipelines ingesting 50+ machine data sources via TCP/IP.
- Cut reporting time by 40% by prototyping a full-stack web app for 24/7 real-time visualization.
- Boosted decision-making by 25% by building Tableau dashboards from 200K+ data points into executive KPIs.
Message me
Short note — I'll reply immediately in hours.