Technical leadership · Production AI systems · Hybrid retrieval · Distributed platforms · Reliability engineering
I build the parts of AI systems that still have to work after the demo: retrieval quality, tool boundaries, stream safety, retry behavior, evaluation loops, human approval gates, observability, infrastructure, and the engineering habits around them.
I work where LLMs meet backend engineering: agent orchestration, RAG, search, evals, governance, distributed systems, platform reliability, and the unglamorous failure paths that decide whether an AI product survives real users.
I am not trying to be a pure people manager. I have led teams, hired engineers, run delivery systems, written RFCs, and handled performance management, but my strongest work is still hands-on technical leadership: shaping systems, debugging hard edges, and helping teams make better engineering decisions.
Building production agentic workflows, hybrid retrieval, human-in-the-loop controls, and cloud infrastructure for enterprise decision support.
| System area | Current focus |
|---|---|
| Agentic architecture | Multi-step agent workflows, tool boundaries, approval gates, typed failure paths, and quality checks before autonomy expands |
| Retrieval & discovery | Hybrid lexical / semantic retrieval, ranking, query interpretation, and graceful degradation for vague or under-specified requests |
| Conversation systems | Stateful conversational flows, streaming behavior, guarded tool use, and product-facing interaction design |
| Platform engineering | AWS-based service architecture, infrastructure-as-code, data stores, security controls, and rollback-friendly deployment paths |
| Reliability hardening | Retry classification, stalled-stream handling, observability, resource isolation, and explicit operating constraints |
I use GitHub less as a trophy case and more as a public engineering notebook: tools, experiments, reliability patches, and production-inspired patterns that expose how I think.
| Repo / work | What it demonstrates |
|---|---|
geehexx |
Canonical GitHub profile and CV source repo with typed YAML adapters, RenderCV/Pandoc generation, deterministic artifact QA, CI-uploaded review packages, and public/resume contact-boundary policy. |
library-ops |
Live Django/PostgreSQL library-operations demo with Work/Edition/Copy modeling, role boundaries, search, circulation state, Render deployment, and a governed multi-agent SDLC case study. |
mcp-web |
MCP server for web and local-file summarization: extraction, chunking, streaming summaries, local/cloud LLM support, caching, security controls, metrics, and tests. |
hitl-mcp-cli |
Human-in-the-loop MCP server and terminal UI for approvals, choices, confirmations, notifications, interaction logging, and async-first workflows. |
phraseturner |
Text-analysis MCP server with personas, readability/naturalness/tone scoring, semantic persona search, and graceful degradation. |
PragmaLens |
Evidence-oriented extraction prototype exploring structured outputs, explicit support spans, uncertainty, and validation-oriented prompt design. |
llm-bedrock-converse |
LLM plugin for AWS Bedrock Converse with streaming, tool calling, MCP integration, multimodal inputs, embeddings, and retry behavior. |
msteams-mcp |
MCP server for Microsoft Teams interaction: message search, replies, thread access, authentication, and assistant-facing collaboration workflows. |
kiro-proxy fork work |
Fork work around streaming, retry behavior, quota classification, model-name preservation, regex safety, and failure-mode cleanup. |
A lot of applied AI engineering is not prompt cleverness. It is removing ambiguity from failure paths.
- Retry storms under fan-out - bounded retry budgets, token buckets, and backoff behavior that does not amplify provider pain.
- Hard quota vs transient throttling - body-content classification so retries do not burn budget against unrecoverable walls.
- Silent stream failures - per-chunk stalled-stream protection and typed terminal errors instead of partial, misleading output.
- Unsafe tool autonomy - approval gates and write-path blocks around irreversible actions, credentials, and protected directories.
- Weak retrieval confidence - hybrid retrieval, rank fusion, query scoping, and evaluation before expanding agent autonomy.
- Operational folklore - ADRs, RFCs, runbooks, and decisions written down before they become archaeology.
| Context | Signal |
|---|---|
| Stealth Startup | Lead AI Engineer for production agentic AI, retrieval, governance workflows, and platform infrastructure |
| BaxEnergy (a Yokogawa Company) | Engineering maturity assessment and modernization roadmap across delivery practices, incident response, DORA metrics, and cloud-native migration planning |
| Agoda (Booking Holdings) | Engineering Manager for B2B booking platforms; led multiple squads and supported high-scale booking flows across millions of properties and major enterprise partners |
| Toptal | Senior Software Engineer, Product Manager, and Engineering Manager across talent matching, vetting, ETL, enterprise delivery, and remote engineering systems |
| Dubizzle (Naspers-backed) | Built and scaled classifieds marketplace systems, rebuilt core ad-placement workflows, and helped move a Django monolith toward service-oriented architecture |
Earlier systems work:
- Independent Freelance / Contract Engagements - Built a greenfield ecommerce platform for an Australian equestrian-event booking startup using Python, Django, Angular 2, payment flows, admin tooling, and deployment ownership.
- Coins.ph - Led Django 1.8 LTS to 1.10 and Python 2.7 to 3.5.2 modernization across the payment-processing platform.
- ITP Media Group - Developed and maintained Django-based CMS platforms powering 40+ websites for the largest publishing house in the Middle East, serving 100+ media brands.
AI / agents / retrieval
- Production agentic AI · RAG / Agentic RAG · hybrid search · vector databases · pgvector · Qdrant · Pinecone · BM25
- ParadeDB · Reciprocal Rank Fusion · PydanticAI · LangGraph · LangChain · AWS Bedrock · MCP / Model Context Protocol
- LLMOps · model evaluation · ONNX Runtime · sentence-transformers · FastEmbed · Optuna · LangSmith
Backend / distributed systems
- Python · Kotlin · Scala · Java · Ruby · Rails · Django · TypeScript · REST APIs · GraphQL · gRPC · PostgreSQL · Redis
- RabbitMQ · Kafka · BigQuery · Avro · Pandas · Elasticsearch · Sidekiq · event-driven architecture · Saga pattern
- high-availability systems · system design
Infrastructure
- AWS · AWS CDK · ECS Fargate · RDS PostgreSQL · ElastiCache Redis · Cognito · WAF · Kubernetes · Istio · Docker
- Terraform · OpenTofu · CI/CD · cdk-nag · observability · platform engineering · rollback-friendly deployment
Technical leadership
- technical roadmaps · architecture decisions · ADRs · RFCs · engineering maturity assessment · DORA metrics
- incident management · on-call design · hiring · headcount planning · mentorship · performance management
- stakeholder management · technical due diligence · remote-first delivery
- Typed errors over mysterious fallthroughs.
- Guardrails around irreversible tool actions.
- Evals before more autonomy.
- Observability before optimism.
- Explicit retry budgets; no infinite faith in exponential backoff.
- Hybrid retrieval when lexical precision and semantic recall both matter.
- Infrastructure that rolls back cleanly.
- Teams that write decisions down before decisions become folklore.
I got here the long way: programming young, learning by building, breaking things, and repeatedly deciding I needed to understand the layer underneath the abstraction.
That meant mIRC scripting, C++, BSD/network programming, GTK+, ncurses, Linux without a GUI for longer than was sensible, Django before it was boring, marketplace systems, fintech workflows, booking platforms, talent-matching systems, and now production AI systems.
The through-line is not a framework. It is systems curiosity plus enough production scar tissue to know where prototypes usually fail.
I am usually useful for:
- building production AI / agent platforms;
- taking RAG systems from prototype to measurable quality;
- hardening distributed systems under real load;
- designing human-in-the-loop governance for AI workflows;
- building MCP and developer workflow tools;
- leading backend / platform teams without drifting away from the code;
- diagnosing reliability failures across infrastructure, code, process, and team boundaries;
- LinkedIn: linkedin.com/in/ancrozier
- GitHub: github.com/geehexx
- Location: Sydney-based Australian citizen
- Open to: Engineering Manager, applied AI, software engineering, backend/platform engineering, and hands-on technical leadership roles. On-site/hybrid across Australia, remote across APAC / EU / UK overlap preferred