Hey, I'm
Deeptimaan
Banerjee
Building scalable ML systems, cloud-native infrastructure, and full-stack applications. Currently ML Team Lead @ Vosyn.
The TL;DR
Machine Learning and Software Engineer specializing in scalable ML systems, cloud-native infrastructure and full-stack development. Proficient in Python, Java, and SQL, with Microsoft Fabric, AWS, and Google certifications. Currently leading ML efforts at Vosyn — previously built privacy-preserving data pipelines and taught 250+ graduate students at the University of Colorado.
My Stack
Languages
ML / AI
Cloud & DevOps
Frameworks & Tools
Where I've Been
- Deployed a self-managed GitLab instance on Google Cloud VM, enabling centralized code management and collaboration across the ML team.
- Built CI/CD pipelines for automated model training and evaluation on language translation pairs with MLflow tracking, streamlining model experimentation and deployment.
- Fine-tuned multilingual ML models (mBART-50, DeepSeek, etc.) for language translation on GCP Vertex AI, building scalable Cloud Run + GCS pipelines and improving translation KPIs by ~15%.
- Built privacy-preserving, real-time data pipelines using a custom WebRTC client and federated learning to continuously train engagement models at scale.
- Fine-tuned and deployed ML/NLP/LLM systems (CNNs, BERT, RAG, LLaMA2) with transfer learning, improving model accuracy and KPIs on secure, highly available AWS/GCP infrastructure.
- Analyzed multimodal behavioral data and led applied ML instruction, driving predictive insights via statistical/causal methods and mentoring 250+ graduate students in advanced AI/ML.
- Automated research insight extraction using NLP/CV (TF-IDF), eliminating ~20 minutes of manual work per product-claim entry.
- Built secure, scalable data platforms replacing Excel workflows with optimized PostgreSQL databases and real-time dashboards.
- Delivered executive analytics dashboards in Power BI and Tableau from Salesforce data across APAC/EMEA to drive sales and product strategy.
Research & Builds
Developed a hybrid CNN–Mixture-of-Experts architecture for facial emotion recognition, achieving SOTA performance — 74.4% ± 0.45 on AffectNet7 and 71.98% ± 0.66 on AffectNet8.
Trained and optimized a transfer-learning CNN achieving 97.24% test accuracy for brain tumor detection from MRI scans with reproducible evaluation pipelines.
Academic Background
Let's Build Something
Whether it's an ML collab, a full-stack project, or just a good tech convo — my inbox is open. Let's connect.