Hi, I'm Kalyan
Machine Learning Engineer | LLM · MLOps · NLP · Computer Vision
About
Results-driven ML Engineer with production experience across the full model lifecycle — data ingestion, fine-tuning, evaluation, and cloud deployment. Adept at parameter-efficient training (LoRA/QLoRA), scalable MLOps pipelines, and cross-functional team delivery. Seeking remote roles building impactful AI systems.
Work History
Machine Learning Intern (AI Developer)
May 2025 – Sept 2025- Sourced and preprocessed 100K+ Indic-language text samples; applied data augmentation and deduplication, raising downstream benchmark accuracy by 15%.
- Designed modular PyTorch training pipelines with reusable data loaders and cross-validation, improving experiment reproducibility across the team.
- Applied LoRA & QLoRA (HuggingFace PEFT) to fine-tune GPT-style decoder models collaboratively with a 5-member team, reducing GPU training time by 20%.
- Packaged and deployed model checkpoints to AWS SageMaker endpoints; logged metrics, hyperparameters, and artefacts in Weights & Biases for full experiment traceability.
My Skills
Check out my latest work
I've worked on a variety of AI/ML projects, from memory systems to workflow automation platforms. Here are a few of my favorites.
Multimodal Sentiment Chatbot
Built a multi-model inference router combining cloud (SageMaker) and local (Phi-2) backends, with dynamic routing based on query complexity. Fine-tuned BERT for sequence classification on 160K labelled tweets; achieved 93% F1-score with voice input/output via STT/TTS.
Crop Disease Detection System
Achieved 96% test accuracy on a Grape Leaf Disease 4 Class using transfer learning (CNN backbone) with augmentation and fine-tuning of classification head. Deployed model behind a REST API (FastAPI) with a Streamlit dashboard for real-time image upload and inference.
I like building things
Innovation 2024 Ideathon (1st Place)
Dec 2024Volunteered as team lead (20+ teams); pitched a drone-based precision agriculture system — onboard camera captured field imagery, a CV model flagged disease zones, and an automated report pinpointed affected coordinates.
National Hackathon (1st Place)
Nov 2023Prototyped and trained an image-classification pipeline end-to-end in 36 hours; ranked 1st among 80+ teams.