Rushikesh – Python, OpenAI, LangChain

An AI platform for the UK energy sector supporting automated document processing, regulatory compliance checks, and agentic chatbots for energy-sector clients. The product combines domain-adapted LLMs, LangGraph agentic workflows, and RAG pipelines to ground model outputs in source data, with production ML infrastructure designed for secure, reliable, and cost-efficient operation.

• Led LLM development for the UK energy sector, including LoRA fine-tuning of LLaMA-7B on domain-specific corpora and reduced the domain-specific error rate by 12% against the base model;
• Designed and shipped specialised LangGraph agentic workflows and tools powering automated document processing, regulatory compliance checks, and agentic chatbots for energy-sector clients;
• Built retrieval-augmented generation pipelines over domain document corpora, covering embedding generation, vector indexing, retrieval tuning, and prompt design to ground model outputs in source data;
• Designed and shipped a comprehensive data ingestion and preprocessing pipeline that standardised training-data collection and cut data-preparation cycle time;
• Served models through FastAPI inference services containerised with Docker and orchestrated on Kubernetes for low-latency, high-throughput production workloads;
• Owned ML deployment on Google Cloud Platform, including Vertex AI, GKE, and Cloud Run, reducing inference and deployment costs by 35% while maintaining production SLAs;
• Implemented evaluation, monitoring, and guardrail practices around LLM outputs to keep production behaviour safe, reliable, and auditable;
• Partnered with product, engineering, and compliance stakeholders to take models from prototype to secure, production-ready systems.

A research project focused on statistical analysis of foundation models, black-box output interpretation, and post-training quantisation of generative models. The work also covers PBR texture generation and model-confidence evaluation for image-generation pipelines, conducted under a formal research partnership with Epic Games, with outputs prepared for submission to top-tier ML and graphics venues.

• Conducted statistical research on foundation models, focusing on black-box output interpretation and post-training quantisation of generative models;
• Specialised in PBR texture generation and model-confidence evaluation for image-generation pipelines under a formal research partnership with Epic Games;
• Produced research outputs prepared for submission to top-tier ML and graphics venues.

An NLP-powered gifting assistant that helped users discover relevant gift ideas through a conversational chat experience. Deployed on Azure and scaled to thousands of users over the course of one year.

• Led product and ML engineering from zero to a live product serving approximately 100,000 users;
• Bootstrapped the startup alongside the Siriolabs.ai role with no external funding.

An intelligent document-processing solution supporting document understanding, extraction, and classification for a Malaysian government workflow serving approximately 3,000 users. The platform used Hugging Face transformer models, including BERT and the LayoutLM family, with production inference services deployed on AWS.

• Led model-optimisation work including pruning, ONNX conversion, and parameter-efficient fine-tuning, reducing model size and improving inference latency on production hardware;
• Ran advanced statistical analysis across large domain-specific datasets using SQL and Pandas to inform feature engineering and model selection.