Niramay Mehta

MSc Artificial Intelligence (Distinction) with hands-on experience building and deploying production AI systems across RAG, multilingual NLP, and generative vision. Delivered end-to-end solutions spanning data ingestion, hybrid retrieval, streaming inference, evaluation, resilient API services, and MLOps using Python, PyTorch, Azure AI, and GCP. Now seeking AI/ML engineering roles focused on hands-on model, data, and platform work.

AI Engineer at Boots UK (2025-05 – Present)

• Developed and deployed a production RAG pipeline for a 100K+ SKU e-commerce catalogue. Built ETL pipelines from Azure Synapse and represented each product as a parent document with linked child chunks for descriptions, ingredients, usage guidance and safety warnings, while retaining identifiers, brand, category, price, stock and restriction attributes as structured metadata.
• Configured hybrid BM25 and vector retrieval within a shared Azure AI Search index, using separate retrieval profiles for product-discovery and pharmacy-information queries. Implemented a GPT-4o-mini intent router, with deterministic downstream rules selecting retrieval weighting, refusal behaviour and response paths; prescription-only medicines were excluded from the searchable index.
• Built a FastAPI query service for real-time streamed generation, achieving P95 time-to-first-token below two seconds in staging load tests from retrieval through to the first generated token. Added dependency-specific resilience through retries, fallback model routing, cached retrieval results and controlled templated responses for unrecoverable failures.
• Integrated Langfuse for request tracing and developed an offline evaluation framework covering faithfulness, context precision and Recall@10. Incorporated evaluation thresholds and stakeholder approval into the Azure DevOps release process for controlled blue-green deployments to AKS.

AI Engineer (Contract) at Ted Baker (2024-05 – 2024-10)

• Architected an end-to-end generative imagery system for internal design and content teams: schema-constrained Gemini prompt expansion, Imagen 3 scene-plate generation, frozen pretrained IDM-VTON garment compositing, and a low-denoise SDXL img2img pass to harmonise lighting, shadows and colour temperature.
• Designed version-aware semantic and content-addressed caching across scene and composite layers, keying reuse to model, checkpoint, prompt-template and asset versions to prevent stale backgrounds or composites after pipeline updates.
• Implemented sequential automated quality gates using masked CIEDE2000 checks for garment-colour fidelity and CLIP ViT-L/14 for prompt-image alignment, with bounded regeneration retries, dead-letter routing and designer review only after automated checks passed.
• Built deterministic Responsible AI controls around structured Gemini outputs, raw-to-expanded prompt diff logging and a rules-based bias gate, backed by tiered tests for schemas, cache invalidation, retry behaviour, prompt contracts and quality checks; instrumented human calibration to identify rule and golden-set blind spots before handoff.

MSc in Artificial Intelligence – Queen Mary University of London (2025-09)

BSc (Hons) in Computer Science – Queen Mary University of London (2024-07)