Headline

AI Engineer | 7x Lower Cost Inference Pipelines | Built LLMs & Diffusion from Scratch

About

I build and optimize AI systems from scratch, for real-world speed, cost, and scale. AI Engineer specializing in Generative AI, LLMs, and hardware-aware optimization. I don't just understand the theory, I ship systems that run fast and cut costs. What I've built: Trained a 1.6B parameter Diffusion Transformer (Flow Matching, DiT) from scratch on 6M images across 5 aspect-ratio buckets on an NVIDIA B200. Pretrained a 160M GPT-2-style LLM on 22B tokens of Cosmopedia at 400k tokens/sec on an H100 Built a complete Stable Diffusion pipeline from scratch, VAE (recon loss → 0.01), CLIP encoder (97% image / 95% text accuracy), cross-attention UNet over 70K steps on LAION Cut inference costs 7x ($2.67→$0.40/hr) via 4-bit quantized 7B model migration Stack: PyTorch · CUDA · FastAPI · Docker · AWS · RunPod Portfolio: dheerentejani.netlify.app GitHub: github.com/dheeren-tejani

Location

Mumbai Metropolitan Region

AI stack

claude

gemini

hugging face

kimi

opencode

RunPod

Work experience

AI Engineering Intern at TheAlly, Jan 2026 - Jun 2026 · 6 mos, Bengaluru, Engineered an end-to-end PDF parsing pipeline, replacing a 70B LLaMA system with a 4-bit quantized 7B Qwen model. Achieved 10x parameter reduction and slashed inference costs by ~7x ($2.67/hr on H100 to $0.40/hr on L4). Developed a FastAPI backend with async job queuing and encryption for structured document processing.

Full Stack Developer at Vivekanand Education Society's College of Arts, Science & Commerce (Autonomous), Jul 2025 - Mar 2026 · 9 mos, Mumbai, Led the end-to-end development of the college's official serverless examination platform. Digitized the remuneration workflow for 100+ faculty members, eliminating manual paper processing and reducing administrative overhead to near zero. Engineered a zero-cost serverless backend using Google Apps Script, ensuring high availability and seamless data integration with Google Sheets APIs. Developed a lightweight, high-performance frontend using Vanilla JavaScript (ES6+) and CSS3, optimizing load times for mobile devices on slow college networks. Managed the full SDLC, from requirement gathering with college administration to final production deployment and maintenance.

Artificial Intelligence Intern at Acmegrade, Apr 2024 - Jun 2024 · 3 mos, Mumbai, Maharashtra, India, Completed an 8-week intensive internship track focused on Computer Vision and Deep Learning. Designed and implemented CNNs for multi-class image classification, achieving high accuracy on benchmark datasets. Optimized model architectures using PyTorch, focusing on hyperparameter tuning to reduce overfitting. Gained hands-on experience in the complete model lifecycle, from data preprocessing to evaluation.

Education

Bachelor's of Science in Artificial Intelligence, Artificial Intelligence, Vivekanand Education Society's College of Arts, Science & Commerce (Autonomous)

Public links

github: https://github.com/dheeren-tejani

dheerentejani.netlify.app: https://dheerentejani.netlify.app/