Curriculum Vitae
Ryan Duffy
AI Infrastructure · Site Reliability Engineering · Applied AI Research
Ipswich, UK · rduffy.uk · blog.rduffy.uk · linkedin.com/in/rduffy-uk
I build and operate AI systems end-to-end. I designed and run Rootweaver — an operator-owned AI platform on K3s that combines typed memory, structural (PDG) code intelligence, signed provenance, and multi-agent verification over an authority-aware vector + SQL + graph retrieval plane, with GPU serving (vLLM) and full-stack observability. AI agents perform substantial implementation work; I own the problem selection, architecture, controls, acceptance criteria, verification and operational outcome. That engineering sits on 14 years of reliability work at BT, currently as an SRE Specialist on Global Fabric — BT International's next-generation global NaaS platform — where I'm embedding Google SRE practice (SLIs, SLOs and error budgets) into a telco operating at global scale.
Selected Research Engineering
- Memory and retrieval evaluation — designed a five-condition pilot across real engineering-decision tasks, locking the rubric before scoring and measuring quality, latency and token cost; results remain preliminary (N=5, one run per cell)
- Faithfulness evaluation — tested an LLM judge across 100 documents and 400 evaluation rows; traced a failed AUC gate to the corruption treatment, redesigned it and reached AUC 0.938 on rerun
- Governed self-improvement — designed offline skill-routing experiments with frozen validation and holdout sets, adversarial cases, matched budgets and independent admission; candidates remain propose-only
Rootweaver — Self-Built AI Platform
AI engineering platform on K3s (RTX 4080, 64GB DDR5) across 54 namespaces. 10 HLD modules, 9 LLD modules, 116 ADRs. 32-package uv workspace.
Explore the live system: architecture decisions · platform evolution · engineering write-ups
vLLM (Qwen3-4B-Instruct-2507-AWQ, 112K ctx) + Ollama CPU fallback
Qdrant ~449K vectors at 2560d, BM25 + vector fusion, cross-encoder reranking, 3-tier agent routing
FalkorDB graph + PDG pipeline, OPA policy enforcement, DORA metrics
Postgres-backed, GDPR tombstoning, 6 recall MCP tools
Kafka → Flink → vLLM workers, claim-check, DLQ replay
Prometheus, Grafana, OTel, Tempo, Loki, Beyla eBPF
149 Sealed Secrets, 227 NetworkPolicies, audit logging
FluxCD, Prefect, KEDA, Harbor registry, Unleash flags
Experience
SRE Specialist
Current- Driving adoption of Google SRE practice — leading SLI/SLO discovery and defining the evidence required before formal targets and error-budget governance
- Observability platform design — crafted architecture, scope and acceptance criteria for AWS-to-on-premise service journeys, including synthetic tests that expose failures earlier
- AI-assisted engineering adoption — championed BT-approved Kiro across an SRE team of approximately 15; more than half now use it daily, with adoption extending to second-line teams and developers
- Production change review — review developer-owned Ansible, YAML, FluxCD and Terraform changes for failure modes, blast radius, observability gaps and rollback requirements
- Alarm correlation — drafted correlation logic across multiple scenarios (PoP isolation, orbiting IS-IS/OSPF links); first scenario now operational
- Recovery and incident engineering — isolated an InfluxDB OOM loop during a 17-hour telemetry incident and validated PostgreSQL WAL/PITR recovery for a 28GB NetBox database, improving tested RPO to approximately one minute and achieving a 26-minute RTO
- BT Tech Fellowship — AI Squad member, contributing to on-premise private cloud AI data centre initiative with NVIDIA advisory
Senior Delivery, Operations & Support Engineer
- Provided second-line engineering across BT's global IP/MPLS and Internet backbone networks, resolving BGP, QoS, bandwidth and customer edge-to-core faults
- Led P1/major incident restoration and RCA for critical infrastructure serving multinational financial services, automotive and logistics customers
- End-to-end support: Cloud Connect (MPLS-to-AWS/Azure), Agile Connect SD-WAN, MPLS core, NFVi, BT Connect
- Multi-vendor fault coordination across international service paths, SLA management and contractual compliance
Delivery, Operations & Support Engineer
24/7 rotational shift cover for BT's MPLS network platforms. Major incident handling and escalation management.
Case Management Analyst
Case management and complaints handling for BT TV/BT Sport. Complex fault diagnosis, root cause analysis, CI/Lean development.
Service Desk Analyst
Priority Care team for BT TV/BT Sport. Complex fault diagnostics, coaching and floor-walking junior colleagues in a supervisory capacity.
Technical Skills
Hypothesis and baseline design · Evaluation rubrics · Failure analysis · Reproducibility · LLM-as-judge calibration · Cost/latency measurement
vLLM · Ollama · Qdrant · FalkorDB · RAG pipelines · Hybrid retrieval · Embeddings · KEDA · MCP · Agent verification
Kubernetes (K3s) · Docker · FluxCD · Helm · HAMi GPU · Harbor · Sealed Secrets
SLO/SLI/SLA frameworks · Error budgets · Burn-rate alerting · RPO/RTO · PostgreSQL WAL/PITR · OPA · Incident management · ITIL
Prometheus · Grafana · OpenTelemetry · Tempo · Loki · Jaeger · Beyla eBPF · Blackbox Exporter · ELK · ServiceNow ITOM
Python (uv-workspace) · Ansible · Prefect · Pydantic · FastAPI · Bash · Git
MPLS · BGP · OSPF · IS-IS · Segment Routing · PCE/PCEP · YANG/OpenConfig · SD-WAN · IPSec
Cisco · Juniper · Nokia · Itential · AWS · Cloudflare · Linux · Jira · Confluence