From semantic embeddings to deployed APIs, I build systems that don't just predict — they perform.
I'm a Software Engineering student (graduating engineer) with a focus on Machine Learning, NLP, and AI system deployment. I've shipped models into production, integrated LLMs into real business workflows, and conducted research at an international institution. I care about the full arc — from problem framing to deployed, documented, maintainable solutions.
- Research intern at Manipal Institute of Technology, India — brain tumor segmentation, self-supervised learning
- Built and integrated a GPT-4-powered assistant into a production application at an enterprise software firm
- Certified by NVIDIA DLI in Transformer-Based NLP, Deep Learning, and Generative AI
- Active in the AI community as a trainer and technical instructor
Languages
ML / AI / NLP
APIs & Deployment
Tools
The Problem: Standard recommender systems rely on a single strategy — either collaborative filtering (who liked what) or content-based filtering (what is it). Each approach fails in isolation: cold-start for collaborative, over-specialization for content-based.
The Solution: A unified REST API that combines both approaches in a hybrid pipeline, routing inference based on data availability and context.
Stack: Python · FastAPI · Collaborative Filtering · Content-Based Filtering · Embeddings
- Production-ready architecture with modular, extensible design supporting real-time inference
- View Repository →
The Problem: Medical image segmentation requires models that are both accurate and interpretable — clinicians need to understand why a region was flagged, not just that it was flagged.
The Solution: A hybrid self-supervised pipeline (ConvMAE + BYOL) trained without full annotation dependency, augmented with Grad-CAM for visual explainability.
Stack: Python · PyTorch · TensorFlow · ConvMAE · BYOL · Grad-CAM
- Achieved a Dice score of 0.94 — state-of-the-art benchmark for this segmentation task
- Grad-CAM overlays provide clinician-interpretable heatmaps, directly addressing model trust in medical AI
The Problem: Keyword-based search returns results that match text, not meaning. A search for "lonely astronaut" won't surface a film tagged only as "space isolation drama."
The Solution: A TF-IDF-based semantic mini search engine that ranks results by contextual relevance rather than raw keyword overlap.
Stack: Python · TF-IDF · NLP · Vectorization
- Demonstrates core information retrieval and semantic similarity concepts that underpin production NLP systems
- View Repository →
| Experience | ~1 year across research and enterprise internships |
| Domains | Medical AI · Enterprise Software · Recommender Systems · NLP |
| Research | International research at MIT India — published pipeline results in brain tumor segmentation |
| Teaching | AI instructor at NATEG Issatso — designed and delivered hands-on training sessions |
| Community | Organizing Committee Co-Lead, Info Plus FST Tunis · AIESEC Global Talent |
I'm open to AI/ML engineering roles, research collaborations, and freelance NLP projects — remote-friendly.
