Neil Cayton

Business Systems Engineer

I build systems that refuse to guess.

Document automation, retrieval chatbots, and event-driven orchestration — in production, with numbers attached. When confidence is low, my systems stop for a human instead of writing a wrong record.

Remote for UK hours · based in the Philippines, GMT+8 · BS Computer Science, Saint Louis University (2026)

Neil Cayton
Book a 30 min call LinkedIn GitHub Email

prompt

reply

Who is Neil Cayton. Neil Cayton, Business Systems Engineer. 5,900 documents to date. 9,900 line items across 11 layouts. Low confidence stops for a human, never guessed. To talk to Neil, use the booking link on this page.

Documents are classified before they are read. Rules handle the layouts the system knows, a model handles the ones it does not, and broken pages are read as images. Anything the system is not sure about stops in one review list with the reason attached — the same gates now generalize to invoices, bank statements and KYC packs.

5,900

documents processed to date

9,900

line items extracted and normalized

11

document layouts in production

1

review queue — low confidence stops here, never guessed into the ledger

Selected work

Four systems, each with the same spine: parse the mess, normalize it, validate before writing, and fail loudly.

01 · Production · current role

Document automation pipeline

The order ledger for a UK food and beverage supplier selling into large national grocery chains runs on a pipeline I own end to end: ingestion, classification, extraction, validation, deployment.

  • Classification before reading — rules for known layouts, a model for the rest, broken pages read as images
  • One review queue with the reason attached; nothing silently guesses into the ledger
  • Grew from the first layout to 11 layouts without adding reviewers
PythonNode.jsLLM APIsPostgreSQLSupabase Edge FunctionsDocker Compose
Case study (PDF) Accuracy & thresholds (PDF)
Documents in email · PDF · scan Classify rules first, model for unknowns, OCR fallback Extract + validate Confidence gate high Ledger grocery POs low Review queue reason attached PostgreSQL state · Slack audit log · rate-limited retries · loud failures

02 · RAG chatbot

Grace — support chatbot that keeps answering

A support assistant that answers from the company's own documents — and keeps answering when an LLM provider goes down, because the provider is treated as an unreliable dependency, not a prerequisite.

  • RASA NLU for intent, FAISS retrieval over the knowledge base, FastAPI backend
  • Multi-provider fallback: an outage degrades to the next model, not an error page
  • Answers cite the retrieved source, so a wrong answer is traceable to a document
PythonFastAPIRASA NLUFAISSMulti-LLM fallback
User message FastAPI backend RASA NLU intent + entities FAISS retrieval, top-k LLM router primary → next Answer + cited source provider outage → automatic failover to the next model

03 · Data pipeline

Order extraction — GPT-4 normalization

Customer purchase orders arrive as email bodies, PDF attachments and scans, in formats that disagree with each other. The pipeline parses each layout, normalizes with GPT-4, and validates against one schema before anything is written.

  • Multi-format parsing: email text, structured PDFs, scanned images via OCR
  • GPT-4 normalization into one schema — quantities, SKUs, dates, delivery windows
  • Validation gate before write; failures return to review with the raw text attached
GPT-4PythonSchema validationOCRETL
POs in email · PDF · scan formats disagree Layout parser per-format rules GPT-4 normalization Schema validation Structured orders JSON failure → review with raw text attached 9,900 line items · 11 layouts · one schema

04 · Infrastructure

Event-driven orchestration

Company operations run through orchestrated workflows: n8n DAGs with custom Node.js triggers, webhooks across half a dozen platforms, and Postgres as the source of truth for execution state.

  • LLM-in-the-loop remediation: when a step fails, a model decides the response and retries under rate limits
  • Webhook integrations: Notion, Microsoft Graph, Slack, and an ERP platform
  • Containerized Python services with health checks and auto-restart via Docker Compose
n8nNode.jsPostgreSQLDocker ComposeWebhooksSlack
Webhooks in Notion · Microsoft Graph Slack · ERP platform Custom triggers Node.js, event-driven n8n DAG orchestration LLM remediation decide → retry, rate-limited Python services Docker · health checks Next step or resolved PostgreSQL — execution state + error logs Slack — every failure, loudly

Earlier work

QA

NOAH — System Integration Testing

QA internship: integration testing across release candidates, reproduction and regression passes before production sign-off.

Mobile

DOST-funded React Native app

Cross-platform mobile build for a government-funded project, from spec to device testing.

Community

AWS Learning Club · SLU

Cloud learning community at Saint Louis University — hands-on sessions alongside the BS Computer Science degree (2026).

Where a wrong record costs more than a delay

The same gates apply to invoices, bank statements and KYC packs. If your documents disagree with each other, that is the exact problem I work on.

Book a 30 minute intro call
neilcayton03@gmail.com LinkedIn GitHub