Aakarshan Bhatia
AI Engineer · London

I build AI products from scratch to launch.

RAG pipelines, AI agents and tool calling, taken from design to deployment. MSc in Artificial Intelligence (Distinction) from Queen Mary, plus four years testing payment infrastructure at FIS Worldpay, so I build with production in mind.

Open to AI engineering roles GitHub LinkedIn CV (PDF)

Ask about my experience

RAG · hybrid retrieval · open model via OpenRouter

    Your question is embedded and matched against chunks of this site, my CV and my GitHub READMEs with hybrid search (vector similarity plus BM25, fused by reciprocal rank). The top six go to an open model that answers only from them and cites each claim; the sources it used are listed below the answer.

    Selected work

    Projects

    Full-stack RAG application

    LectureMate

    A study assistant for lecture PDFs and slide decks. Students upload course material, ask questions, and get answers grounded only in what they uploaded, with citations like Week 3 Slides, page 12. If the answer isn't in the material, it says so.

    Model providers are swappable: Ollama runs locally by default (llama3.1, nomic-embed-text), and OpenAI is one environment variable away.

    FastAPIReact + TypeScriptWeaviateOllamaOpenAIAWS S3Docker
    View code on GitHub
    Request path
    1. Upload PDF or PPTX · original stored in S3
    2. Extract text per page and slide
    3. Split into overlapping chunks
    4. Embed and index · Weaviate
    5. Retrieve top chunks for the question
    6. Answer from retrieved context only · with citations
    MSc dissertation · Queen Mary University of London

    Mental health risk assessment from social media text

    A coarse-to-fine classification pipeline that first separates at-risk posts, then grades severity across imbalanced classes. TF-IDF features with calibrated logistic regression, so the probability scores can be trusted when setting triage thresholds.

    Gains were tested with a paired bootstrap rather than reported from a single split. I also assessed ethical risks, including stigmatisation from false positives.

    +0.075macro-F1 over single-stage baseline
    p < 0.01paired bootstrap
    scikit-learnTF-IDFLogistic regressionProbability calibration
    Recall at 90% precision

    At a fixed 90% precision, the dual-stage model catches 24% more at-risk posts.

    Machine learning app

    Credit risk prediction

    End-to-end model on the German Credit dataset, from EDA to a Flask app that returns a default probability and risk band. Treated "bad" risk as the positive class, used balanced class weights, and shipped preprocessing inside the saved model so training and inference can't drift.

    scikit-learnRandom forestpandasFlask
    View code on GitHub
    Client work · Adaptive Labs

    Agents, tools and MCP

    At Adaptive Labs I build LLM applications that call tools, query vector stores and connect to business systems. I prototype agent-to-tool connectivity with the Model Context Protocol and work across OpenAI, Claude, Gemini and OpenRouter.

    AgentsTool callingMCPEmbeddingsFlutterPostgreSQL
    Career

    Experience

    Sep 2025 – now

    Founder and AI Engineer

    Adaptive Labs · London (hybrid)

    • Founded a software and AI firm; lead design and delivery of production AI applications.
    • Design and build LLM systems with RAG, AI agents, tool calling and agentic orchestration, covering embeddings, semantic search, prompt engineering and model evaluation.
    • Integrate LLMs with external APIs, databases, vector stores and business systems; prototype agent-to-tool connectivity with MCP.
    • Build APIs and full-stack web and mobile apps, from technical design to deployment.
    2020 – 2024

    OAT Test Analyst

    FIS Worldpay · London

    • Led operational acceptance testing across 30+ enterprise fintech projects for critical payment infrastructure.
    • Improved testing efficiency by 30% by introducing Selenium automation and streamlining test processes.
    • Ran an enterprise project end to end, introducing new operating methods that improved efficiency 4x.
    • Designed risk-based test strategies for cloud and on-premise systems across test, pre-production and production.
    • Worked with architects and engineering teams to validate resilience and deployment readiness, catching critical defects before release. Mentored new analysts.
    Toolkit

    Skills

    Generative AI

    LLMs, RAG, AI agents, multi-agent workflows, tool and function calling, MCP, prompt engineering, embeddings, vector and semantic search

    Models and search

    OpenAI, Claude, Gemini, OpenRouter, Ollama, Weaviate, Elasticsearch

    Machine learning

    scikit-learn, pandas, NumPy, NLP, classification, regression, clustering, anomaly detection, feature engineering, calibration, hyperparameter tuning, model evaluation

    Languages

    Python, SQL, TypeScript, JavaScript, HTML, CSS

    Backend, web and mobile

    FastAPI, Flask, Node.js, REST APIs, React, Flutter, PostgreSQL, Firebase

    Cloud and delivery

    AWS S3, Docker, Git, Selenium, test automation, risk-based testing, Agile, SAFe

    Education

    Study

    MSc Artificial Intelligence (NLP), Distinction

    Queen Mary University of London · 2024 – 2025

    BSc Computer Science, First-Class Honours

    University of Hertfordshire

    Software Engineering Virtual Experience

    Goldman Sachs · 2024

    Contact

    Hiring for an AI role? Let's talk.

    LinkedIn GitHub CV (PDF)