Tejas Bana

Agentic Systems for Scientific R&D

Machine Learning Engineer II at Albert Invent

Promoted to ML Engineer II after leading the development of two core agentic systems:

• Agentic Interface: Built a production agent to simplify a complex UI, enabling LLMs to translate user natural language into concrete UI actions and data retrieval. Achieved 91% reliability with RAG and advanced prompting.
• LLM Discovery: Created a chat-based search and recommendation engine scaling to millions of data points, dramatically improving data accessibility for scientists.

Learning-Based Image Synthesis (16-726)

Carnegie Mellon University

A deep dive into generative vision through hands-on projects:

• Main Project: GANs to Understand the Brain
• GAN Photo Editing
• Neural Style Transfer
• When Cats Meet GANs
• Gradient Domain Fusion
• Colorizing Photo Collections

Generative Vision Seminar (16-895)

Carnegie Mellon University

A seminar course exploring the nuances of modern generative models, from understanding their success to critiquing their capabilities for future research. Covered topics like Diffusion Models, VAEs, and Flow-based models.

GANs for Architectural Floorplan Generation

ML Intern at Uniworks Designs

Automated the generation of apartment floorplans and furniture layouts using a two-stage GAN pipeline (House-GAN and ArchiGAN). Deployed the model as a service on GCP, reducing initial design time for architects by up to 25%.

Generative Camouflage for the Indian Army

R&D ML Intern at College of Military Engineering

Developed software to generate pixelated, terrain-specific camouflage patterns. The system uses satellite imagery, segments terrain features with Mask R-CNN, and applies a generative algorithm to create 5K resolution patterns in under 60 seconds, significantly reducing vehicle visibility. Received "Best Intern" award for this work.

Siamese Networks for Fashion Similarity

R&D ML Intern at IIT Bombay

Built a deep learning system to find visually similar clothing items from a massive dataset. Implemented a Siamese network with triplet loss and multi-class negative sampling to learn powerful image representations. Used Annoy for efficient, real-time nearest-neighbor search.