Designing and optimizing systems for large language models — focusing on inference performance, memory efficiency, scalability, and real-world deployment constraints.
Computer Engineering student with a strong focus on ML Systems Engineering, LLM inference optimization, and data pipeline design.
Experienced in building end-to-end systems around large language models, including:
- Transformer inference optimization (KV caching, memory management)
- Large-scale dataset engineering for LLM fine-tuning
- LLM-powered applications (RAG, NL → SQL systems)
Interested in solving problems related to latency, memory constraints, throughput, and scalability in modern AI systems.
ML Systems & Infrastructure
- Transformer inference optimization (KV cache, paging, prefix reuse)
- Memory-efficient model serving and execution
- Performance profiling (latency, memory, compute trade-offs)
LLM Engineering
- Fine-tuning workflows (instruction tuning, dataset preparation)
- Retrieval-Augmented Generation (RAG)
- Prompt engineering and structured generation
Data Engineering for ML
- Large-scale dataset cleaning, deduplication, and normalization
- JSONL dataset design for training pipelines
- Data validation and reproducibility
Backend & Systems
- API development (FastAPI)
- Pipeline orchestration concepts (Airflow, MLflow, DVC)
- Dockerized workflows and reproducible environments
Systems-level implementation of memory-efficient KV caching for transformer inference.
- Designed paging-based KV cache architecture decoupling logical token positions from physical memory
- Implemented prefix reuse, reference counting, and copy-on-write semantics for efficient memory sharing
- Reduced redundant allocations and improved memory efficiency in long-context decoding scenarios
- Validated numerical correctness against naive attention (<0.1% deviation)
Keywords: Transformer Inference, KV Cache, Memory Optimization, LLM Systems, PyTorch
Large-scale data engineering system for low-resource LLM fine-tuning.
- Processed 16M+ tokens (1.1M+ sentences) with automated deduplication and noise filtering
- Built structured JSONL datasets for instruction tuning and evaluation
- Designed reproducible pipelines with schema validation and metadata annotation
- Improved data quality and reduced manual curation effort significantly
Keywords: Data Pipelines, NLP, LLM Fine-Tuning, Dataset Engineering
LLM-powered system for querying structured databases using natural language.
- Built end-to-end pipeline: user query → SQL generation → execution → formatted response
- Integrated with Slack using FastAPI and LangChain
- Implemented schema-aware SQL generation for accurate query execution
Keywords: LLM Applications, LangChain, SQL, Backend Systems, API Integration
- System-first mindset: design architecture before implementation
- Focus on performance metrics: memory usage, latency, throughput
- Explicit analysis of trade-offs (memory vs compute vs scalability)
- Emphasis on reproducibility and pipeline reliability
- Preference for understanding underlying systems over relying only on frameworks
Languages: Python, C++, JavaScript
ML/AI: PyTorch, Transformers, Scikit-learn
LLM Tools: Hugging Face, LangChain, OpenAI SDK
Data: Pandas, NumPy, SQL (PostgreSQL, SQLite)
Backend: FastAPI
Systems & Infra: Docker, Linux, MLflow, DVC, Airflow
Tools: Git, Jupyter Notebook, VS Code
NLP Research Intern — Thapar Institute of Engineering & Technology (2025–Present)
- Built end-to-end data engineering pipelines for Punjabi LLM fine-tuning
- Processed 16.2M+ tokens with automated deduplication, normalization, and validation
- Developed structured datasets for instruction tuning and evaluation workflows
- Improved dataset quality and reproducibility for downstream training systems
- LLM inference optimization (KV caching, memory-efficient decoding)
- Scalable ML systems and model serving architectures
- Dataset engineering for low-resource languages
- System-level aspects of AI safety, robustness, and reliability