Angelos – AI deployment, Python, Vector Databases

An AI research assistant for a business intelligence platform used by corporate, financial and government clients. Subscribers ask questions in plain language over a large archive of country, industry and economic analysis. The assistant answers across several turns, cites its sources, and only draws on content each client has paid for. Its features were agentic multi-turn Q&A, retrieval over 2M+ documents, entitlement-aware search, citations, and access to the content through MCP for internal tools and paying API customers.

• Led R&D on the platform’s first customer-facing AI agent, a multi-turn Q&A product now used by B2B clients including S&P 100 companies.
• Coordinated delivery across five teams: AI platform, data science, engineering, editorial/product and UI.
• Grew engagement by 10% month on month and kept running costs negligible at that scale.
• Designed the RAG architecture on Bedrock and OpenSearch. This covered chunking strategy, recency decay in ranking, and per-client entitlement filtering, so users only see content they’re licensed for.
• Moved the product from search-and-summarise to an agent-based design, and tested GraphRAG approaches along the way.
• Took MCP from proof of concept to production: content store connector, gateways, auth and semantic search tools. It’s now used by internal teams and to sell data access to external customers.
• Built a content recommender over 2M+ documents, from ingestion and embeddings through to vector search and a production API.
• Advised an external partner (Infosys) on a large-scale research agent for building nuclear indices.

The AI platform and GenAI features behind a global news publication’s digital products: article recommendations, automatic tagging and enrichment of the archive, and editorial AI tools. Readers get better content discovery. Editorial and product teams get a shared, self-serve set of AI services (Q&A, summarisation, embeddings, translation) that they can build on without starting from scratch.

• Helped build the group’s AI platform from scratch and set up AWS Bedrock as the shared foundation for Q&A, summarisation, embeddings and translation.
• Migrated GenAI workloads off a third-party platform (Deepset) and retired it, saving £120K a year.
• Rebuilt the article recommender end to end over a 250,000-article archive.
• Built enrichment pipelines with data science that run 20+ tagging processes over 250,000+ articles. They produced 5M+ tags across the archive and now tag new content as it’s published.
• Owned evaluation of the GenAI tools so their output met the publication’s standards for voice and quality.
• Wrote the technical case for using GenAI in an editorial business (evaluation, guardrails, citations) during an AWS Experience-Based Acceleration programme.
• Built an internal AI platform UI so anyone in the company could try the tools, and ran AI workshops for staff.
• Founding member of the internal AI Lab, which prototypes future products.
• Represented the company on agentic AI at AWS Summit London (main stage), London Tech Week and AI Summit London.

An automated online interview platform for selecting candidates, mainly for places in the health professions. Candidates record answers to interview questions in their own time. The platform transcribes them and scores them with AI models designed to keep bias out. Recruiters and admissions teams get structured, comparable results without running every interview live. The main features were asynchronous video interviews, automatic transcription, AI-assisted scoring with fairness principles built in, accessibility support, and dashboards for hiring and admissions teams.

• Was part of the founding team. Led engineering and architecture for both of the company’s core products, from planning through build, deployment and monitoring.
• Built the platform serverless on AWS and had it reviewed against the Well-Architected Framework. It used conversational AI, generative AI, batch transcription and model fine-tuning.
• Built ML pipelines that collect interview text in batches and automatically fine-tune and redeploy the scoring and transcription models. Preprocessing, training and deployment all ran as SageMaker jobs.
• Set up CI/CD for test, production and feature builds.
• Fine-tuned a speech recognition model on SageMaker, cutting transcription costs by 80%. This was part of a wider cost review that reduced total cloud spend by over 75%.
• Grew the platform to 10,000+ users across 11 enterprise customers, who cut interview admin time and costs by over 50%.
• Led an InnovateUK-funded accessibility work package.
• Took the company through technology compliance, which supported partnerships with Airbus, the UN and the NHS, plus 10+ grants, awards and investments.
• Managed customer relationships and SLAs, and worked with non-technical stakeholders and external partners.
• Co-authored two peer-reviewed papers in BMJ Open (2022, 2023) on the method and its fairness evaluation.

A company data and risk intelligence pipeline that builds structured profiles of private and public companies from public web data. It feeds research analysts, risk models and ESG portfolio analysis. The system crawls company websites, newsrooms and PDF reports, pulls out key facts with ML and NLP models, classifies each company by industry, and spots key events in the news. The results are served as APIs that analysts and downstream products use to assess company risk. The main features were large-scale web and PDF extraction, ML-based field extraction (founding year, employee count, headquarters, contacts), industry classification, news event detection, company risk scoring APIs, and ESG portfolio analysis.

• Worked as solutions architect and ML engineer on a three-way partnership, with responsibility for building and deploying the data and ML side.
• Built scrapers for company websites, newsrooms and PDFs, covering 15+ data points per company for up to 100,000 companies.
• Built parsers for founding year, employee count, headquarters and contact details. They passed the extracted text through Hugging Face question-answering models and kept the highest-scoring answer for each field.
• Tuned each parser one field at a time against the client’s own validation dataset to improve accuracy.
• Scraped 10,000 company websites to build an industry classification (NAICS) dataset, and built the training set for a URL classification model.
• Built custom SageMaker models to identify key events in news articles.
• Built a company industry classifier (Spark on SageMaker, Athena, Glue) that retrains automatically from live performance data.
• Built and deployed ML inference APIs that classify and score company risk. They served over 1 million API calls and brought in recurring monthly revenue.
• Cut inference running costs by 70% with caching and reworked SageMaker pipelines.
• Built ESG portfolio analysis and a company description API for the partner’s customers.
• Set up governed deployment: separate build and deploy repositories, a client approval step that triggers production ML pipelines, and full model lineage (data, algorithms, metrics) for audit.
• Deployed all services as Docker containers defined in AWS CDK.

A live construction market insights tool. Cost managers and clients use it to forecast project budgets and decide when to buy materials. Every day it collects raw material prices from UK merchants (timber, steel and others), ONS statistics and global news, and shows an up-to-date market view alongside ML forecasts of where costs are heading. The main features were daily price tracking across 10,000+ items, budget forecasting, purchase-timing recommendations and a live dashboard of market data.

• Led engineering and architecture as the client’s lead engineer and solutions architect, working alongside their data team.
• Built and ran the scraping engine on AWS, which tracks 10,000+ items a day from UK merchants, ONS data and news feeds.
• Processed the scraped data with Spark, orchestrated by Step Functions and Lambda, with everything packaged in Docker so it moves easily between cloud environments.
• Set up the ML platform for the data scientists (SageMaker domains and experiments), then automated model training and deployment to SageMaker endpoints.
• Added automatic data and model drift monitoring against training baselines, with alerts when drift is detected.
• The tool helped the client cut construction costs by 5–10% by timing material purchases.
• It also made the move from planning to the build phase 5x faster.