8 min
Part 7: How to Test a System That Answers Differently Every Time
assertEquals dies on contact with an LLM. Here is what we asserted instead: schema conformance, retrieval provenance, guardrails and adversarial UAT.
Senior AI Automation Engineer
Anh
Trinh
Full-Stack Software Engineer
Sole engineer behind an Australian staffing firm's automation programme — event-driven AWS pipelines with Claude as the reasoning layer, built on the tools the business already runs on. Backed by 7 years shipping production web, mobile and cloud systems across four countries' clients.
Open to new roles location_on Sydney, NSW verified_user Full working rights
workspace_premium
Western Sydney Uni
M.ICT Distinction
smart_toy
AI in production
Claude · Bedrock · RAG
cloud_done
Lambda · SQS · DynamoDB
Event-driven AWS
Currently
Sole engineer on an automation programme in Sydney
Senior AI Automation Engineer at Outsourced Staff Pty Ltd, working directly with the founder. The brief is unusual and I like it: take a description of how a team already works and turn it into automated flows on the tools they already have — no process change, nothing new for anyone to learn.
That means owning each initiative end to end — scope and estimate, architecture, build, testing, rollout, and the runbook someone else can operate it from.
HR Tech
arrow_forwardCandidate Automation
An AWS serverless pipeline that turns Lever ATS activity into client-ready candidate scorecards, with Claude as the assessment layer. Removes hours of manual scorecard preparation per role while keeping a human in front of every client-facing document.
TypeScript AWS Lambda SQS + DLQ DynamoDB EventBridge Claude API
HR Tech
arrow_forwardEmail Automation
A config-driven HR and client communications engine on Make.com, Google Workspace, Reply.io and the WhatsApp Business API. The Admin can reword an email, change its timing or recipients, switch it off, or add a whole new message type by adding one row to a spreadsheet — no developer involved.
Make.com Google Workspace APIs Reply.io WhatsApp Business API Config-Driven Design PDF Generation
500k+ requests load-testedHighly available AWS architecture — no disruption, no data loss
60+ brands on one platformMulti-tenant commerce platform maintained without cross-tenant impact
3 time zones, asyncUS product owner, European engineers, delivery from Vietnam
6.438/7 Master of ICTWith Distinction, Western Sydney University — nine HD subjects
The stack, in two layers
hover any badge for detailAI & Automation
smart_toy
Claude / Anthropic API neurology
AWS Bedrock search_insights
RAG travel_explore
OpenSearch bolt
AWS Lambda queue
SQS + DLQ table
DynamoDB schedule_send
EventBridge deployed_code
Serverless Framework account_tree
Make.com workspaces
Google Workspace APIs forum
WhatsApp & Reply.io smart_toy
Claude / Anthropic API neurology
AWS Bedrock search_insights
RAG travel_explore
OpenSearch bolt
AWS Lambda queue
SQS + DLQ table
DynamoDB schedule_send
EventBridge deployed_code
Serverless Framework account_tree
Make.com workspaces
Google Workspace APIs forum
WhatsApp & Reply.io Full-Stack Foundation
code
TypeScript widgets
React & Next.js change_history
Angular sync_alt
RxJS & NgRx javascript
Node.js coffee
Java & Spring Boot php
PHP Laravel database
PostgreSQL & SQL cloud
AWS Cloud smartphone
Kotlin & Swift package_2
Docker & CI/CD brush
Tailwind CSS code
TypeScript widgets
React & Next.js change_history
Angular sync_alt
RxJS & NgRx javascript
Node.js coffee
Java & Spring Boot php
PHP Laravel database
PostgreSQL & SQL cloud
AWS Cloud smartphone
Kotlin & Swift package_2
Docker & CI/CD brush
Tailwind CSS How I build
Systems that survive the second run
7 years of production delivery, most recently owning an automation programme end to end — scope, estimate, build, rollout and the support afterwards. These are the six things I bring to that, and the evidence behind each one.
The full background arrow_forwardsmart_toy
AI That Ships
LLMs behind a contract, not a demo. Forced tool-calling schemas, confidence gates, retrieval grounded in the customer’s own documents, and a human in front of anything a client sees.
Claude · AWS Bedrock · RAG on OpenSearch
hub
Event-Driven by Default
Webhooks, queues and schedulers instead of cron jobs and hope. Work is decoupled behind SQS with a dead-letter queue, so a downstream outage delays a message rather than losing it.
Lambda · API Gateway · SQS/DLQ · EventBridge
restart_alt
Idempotent Under Retry
At-least-once delivery means every handler will run twice eventually. Provisioning locks, property-keyed upserts and checksum change detection make the second run a no-op.
DynamoDB conditional writes · content hashing
lock
Security by Design
Signature verification on every inbound webhook, least-privilege roles per function, secrets that never reach source control, and tenant data isolated in the query layer rather than the prompt.
HMAC · OAuth · least-privilege IAM · KMS
monitoring
Testing & Observability
Unit and integration suites with in-memory fakes, plus live end-to-end runs against the real third-party APIs. Structured JSON logs you can build a metric filter on, and alerts on the DLQ.
Contract tests · JSON logging · DLQ alerting
swap_horiz
Vendor-Agnostic Cores
Third-party systems live behind an adapter contract, so swapping an ATS, a CRM or a mail provider is a new adapter rather than a rewrite. Proven on a 60+ brand multi-tenant platform.
Adapter pattern · DTO boundaries · Clean Architecture
Writing
45 write-ups across 13 series
Long-form engineering notes on AI automation, AWS architecture, cloud security and the decisions behind them — including a seven-part teardown of a production RAG agent.
8 min
Part 6: Idempotency and Dead-Letter Queues in an LLM Pipeline
In an LLM pipeline a duplicate email is a billing event, so we keyed idempotency to the Message-ID and let a dead-letter queue hold the rest.
9 min
Part 5: The AI Agent That Physically Cannot Send Email
Amazon Bedrock Guardrails mask PII before inference, and an agent with no send capability turns the Drafts folder into a zero-UI approval queue.
Get in touch
Let's automate
the boring part
If you have a process someone is doing by hand, a pipeline that keeps double-processing under retry, or an LLM feature that needs to be safe enough to put in front of a client — that is the work I want.
Open to senior AI automation and full-stack roles in Sydney
- Based in
- Sydney, NSW, Australia
- Work rights
- Full working rights — no sponsorship required
- Open to
- Contract or permanent · hybrid or remote
- Reply time
- Within 24 hours, AEST business days