Ayush Kumar

AI engineer (MSc Advanced Computer Science, Sheffield) who turns workflow pain points into production-ready Python services.

Built and deployed PickR solo: a FastAPI service with a LangGraph agent, vector database, schema validation and monitoring. Builds in controls, from validation checks to human approval.

Right to work: UK Graduate visa applied for, no sponsorship required

MSc Dissertation Researcher at UNIVERSITY OF SHEFFIELD (2026 – 2026)

Fine-tuning vs Retrieval for an LLM (Llama 3.2 3B)

• Built the data pipeline that converted NutriBench meal records into about 5,000 question and answer training pairs, then fine-tuned Llama 3.2 3B with QLoRA on a single A100.
• Evaluated two approaches, fine-tuning and retrieval, with a held-out test set, accuracy scored against EU labelling tolerance bands, and five-fold stratified cross-validation on 5,300+ meals.
• The fine-tuned model reached 69.1% carbohydrate accuracy against 66.8% for GPT-4o. Retrieval made it worse, so the report documented the trade-offs and recommended the simpler, cheaper design.

MSc in Advanced Computer Science – UNIVERSITY OF SHEFFIELD (2025 – 2026)

B.E. in Computer Science – CHANDIGARH UNIVERSITY (2021 – 2025)