Skip to content

Projects

Selected professional projects

Anonymised project evidence from professional product and delivery environments.

My responsibility

Technical design and end-to-end delivery of the integrated AI features — from document processing and retrieval to integration with existing product workflows and permission models.

Problem

In document repositories that have grown over many years, manual classification, metadata maintenance, and domain research create substantial effort. At the same time, AI-powered search must not bypass the existing permission model.

Solution

Integrated AI features classify documents, extract metadata, generate summaries across languages, and answer questions about individual documents as well as authorised document collections.

Result

Document-related workflows were embedded directly into the existing ECM platform rather than introduced as a separate AI tool alongside the product.

Integrated document-AI flow Documents are processed, classified, and retrieved for grounded answers only within the existing permission model. Documents existing ECM collection Process classification · metadata Summarise including cross-language Retrieve with access permissions before selection Answer with evidence accessible collection only Integrated into existing product workflows not a separate AI tool alongside the ECM platform
  1. Documents existing ECM collection
  2. Process classification · metadata
  3. Summarise including cross-language
  4. Retrieve with access permissions before selection
  5. Answer with evidence accessible collection only
  6. Integrated into existing product workflows not a separate AI tool alongside the ECM platform
Technical details and decisions

Technical Architecture

LLM-based document processing and semantic retrieval within existing product workflows. Permissions are respected during retrieval, and answers remain grounded in the accessible document collection.

Engineering Decisions

  • Integrate AI features into existing product workflows
  • Apply permissions during candidate retrieval
  • Treat classification, extraction, and summarisation as verifiable features

Demonstrates

Applied LLM engineering, document analysis, semantic retrieval, RAG-oriented knowledge access, and integration into an enterprise platform.

Status

Anonymised professional project; client and product names are not disclosed.

Large Language ModelsSemantic RetrievalDocument AIEnterprise Integration

My responsibility

End-to-end responsibility for intake, structured extraction, tenant-isolated retrieval, state graph, privacy boundaries, and integration into the web application.

Problem

Incoming cases from multiple channels were triaged manually, domain answers required research across extensive document collections, and deadlines were difficult to manage reliably.

Solution

An idempotent intake pipeline combines classification and structured extraction with tenant-isolated retrieval, source citations, and an explicit state graph. Personal data is processed locally; external models receive masked domain questions only.

Result

The initial domain response was reduced from several working days to a few hours. No case triggered an action without human approval.

AI-assisted case-processing flow Incoming items are structured, grounded in tenant-isolated documents, processed through a state graph, and executed only after human approval. Incoming items multiple channels Structure classification · schema Retrieve evidence tenant · source · page Prepare workflow persisted state graph Human approval no autonomous action Process PII locally external model sees masked domain questions only
  1. Incoming items multiple channels
  2. Structure classification · schema
  3. Retrieve evidence tenant · source · page
  4. Prepare workflow persisted state graph
  5. Human approval no autonomous action
  6. Process PII locally external model sees masked domain questions only
Technical details and decisions

Technical Architecture

Next.js frontend, FastAPI services, PostgreSQL with row-level security, hybrid document retrieval with reranking, and LangGraph as a deterministic state graph persisted after every step.

Engineering Decisions

  • Tenant isolation at database level rather than only in the application
  • Source reference and page number for every domain answer
  • Explicit states and human approval instead of an open-ended agent loop

Demonstrates

End-to-end AI product engineering, RAG, structured extraction, LangGraph workflows, privacy, and human-in-the-loop automation.

Status

Anonymised professional project; client data and identifying details are not disclosed.

Next.jsFastAPIPostgreSQL RLSLangGraphHybrid RetrievalRerankingLocal Speech-to-Text

A relevant role or product challenge?

I am open to permanent senior roles in AI product engineering and to fullstack or frontend roles with an AI focus.

Get in touch