AI Engineer — Agentic AI & Document Intelligence

About the role

We are building agentic AI systems that process real financial documents end to end — scanned documents in, validated and correctly coded records out, exported automatically to downstream business systems, with humans reviewing only what the system is unsure about.

This is not a research role and it is not a chatbot role. You will build pipelines where accuracy is measured, failures are visible, and a wrong output has a real financial consequence downstream. If you enjoy problems where "it worked in my notebook" is the beginning rather than the end, this will suit you.

The work spans AI engineering, data engineering and integration. You will not be handed a narrow slice.

What you will work on

Document intelligence

• Extraction pipelines over multi-page scanned batches — classifying pages, grouping them into logical transactions, and pairing related documents before extraction
• OCR, computer vision and multimodal models combined to read printed and handwritten fields
• Field-level confidence scoring, cross-validation between fields, and handling poor-quality scans honestly rather than guessing

Agents and decisioning

• Multi-agent workflows: an orchestrator coordinating specialised agents for ingestion, data quality, identity resolution and reporting
• Classification and inference agents that fill missing values from business rules combined with historical context
• Deterministic rules engines working alongside model-based judgement, with the model never computing financial values
• Anomaly and outlier detection to flag transactions that need human attention

Data

• Canonical data models in PostgreSQL, normalising records arriving from very different sources
• Entity resolution and deduplication — matching incoming records against existing master data
• Ingestion from CSV, Excel and Google Sheets, alongside API sources
• Export into strictly defined external formats

Integration

• REST API integrations with CRM, payment and accounting platforms
• Payment gateway integrations — transaction retrieval, webhooks for live events, OAuth2 flows
• Secure per-customer credential storage, encryption at rest, and audit logging

Product surface

• Backend APIs serving review and approval interfaces
• Wiring existing HTML prototypes into working, data-driven screens
• Role-based access control across multiple user types, in a multi-tenant environment

Quality

• Human-in-the-loop review flows — confidence thresholds, approval queues, and capturing every correction as labelled data
• Evaluation harnesses that tell us, with numbers, whether the system is getting better or worse
• Cloud deployment on Azure or AWS

Required skills and experience

• 1–3 years total professional experience. This is a firm range covering your entire career, not your AI experience alone — we are hiring at early-career level and applications above 3 years total will not be shortlisted
• Python as your primary programming language
• Hands-on experience building at least one agentic AI system. Not a tutorial or a course project. You will be asked to describe it in detail at interview: what it did, how it was structured, what went wrong, and how you knew it was working
• Practical experience with at least one agent framework: LangGraph, LangChain or CrewAI
• Direct experience with LLM APIs — prompting, function and tool calling, structured output, streaming
• Strong data handling: pandas, SQL (PostgreSQL preferred), data cleaning, validation and transformation across formats
• Consuming and building REST APIs; comfortable with JSON, authentication, pagination, rate limits and webhooks
• Git, and the habit of writing code others can read

Strongly preferred

• Microsoft AI stack: Azure AI Foundry / Microsoft Foundry (Agent Service), Semantic Kernel, AutoGen or Microsoft Agent Framework
• Document AI: Azure AI Document Intelligence, AWS Textract or Google Document AI — any production OCR or form-extraction work
• Azure OpenAI Service, AWS Bedrock or Azure AI Search
• Evaluation of LLM systems — building test sets, measuring accuracy, catching regressions
• Structured output enforcement using Pydantic or JSON Schema
• Entity resolution, record matching or deduplication at scale
• Anomaly or outlier detection — rules-based or statistical
• Retrieval-augmented generation with a production vector store, and judgement about when RAG is the wrong approach
• FastAPI, or similar Python API frameworks
• Cloud deployment on Azure or AWS, containers, CI/CD

Nice to have

• Third-party platform integrations: Salesforce, PayPal, Stripe, QuickBooks, or comparable CRM, payment or accounting APIs
• Google Sheets and Google Forms APIs
• OAuth2 implementation, secrets management, encryption at rest
• Multi-tenant application design and role-based access control
• Frontend capability — React, Next.js, or comfort taking an HTML prototype through to a working screen
• Exposure to financial or accounting data, and an appreciation of why correctness matters differently there
• Async Python, queues and background workers
• Prompt versioning or experiment tracking
• Handling PII responsibly, and awareness of audit and compliance requirements
• Open-source contributions, or a project we can look at

Work Location: In person