AI Systems · LLM Training · Efficient Inference
I build language models from scratch, reproduce modern LLM research, optimize inference systems, and explore efficient AI architectures.
Fast Memory-Efficient Inference Engine for Large MoE Models
Turbo-LLM is an experimental inference engine designed to run very large Mixture-of-Experts language models under strict VRAM constraints using dynamic expert execution and adaptive memory management.
Training a 17.2M Parameter Language Model From Scratch
TinyStories-17M is a decoder-only Transformer trained entirely from scratch on 2.1 million TinyStories-style synthetic stories to explore how capable modern Small Language Models can become through high-quality data and efficient architecture.
From Classical Machine Learning to Knowledge Distillation
An end-to-end email classification system that explores multiple machine learning paradigms, progressing from TF-IDF baselines to deep neural networks and finally a lightweight distilled model for efficient deployment.
Domain Adaptation of Open Language Models for Indian Legal Reasoning
A series of domain-specific fine-tuning experiments exploring how high-quality legal instruction datasets improve small language models for Indian legal question answering, structured reasoning, and long-form explanation.
Structured directory of core technical capabilities. Highlighted areas indicate primary engineering focus in training pipelines and low-latency inference systems.
Problems I return to weekly — an evolving intersection of systems and post-training.
Full research pageSSD → RAM → VRAM expert offloading with double-buffered scheduling and warm-cache routing bias.
Data-density scaling for models under 100M parameters — how far can curated synthetic data go.
Provision-summary formats that stop compact SFT models from hallucinating source-adjacent facts.