Operating Agentic Systems in Production: Lessons from Building Tendwell
The hard part is not the model, it is the guardrails around it - the propose/approve/execute separation, local-first as a hard default, and audit as a feature.
Platform engineering · Regulated environments
Governed AI infrastructure for regulated environments
I build the production platforms that run LLMs and agents where audit, security, and reliability are not optional - for banks, fintechs, and high-stakes operators.
llm-gateway — production
governed · audited
$ curl -s https://llm.internal/v1/chat \
-H "Authorization: Bearer $TEAM_TOKEN"
route: bedrock/claude ALLOWED
audit: request logged TRACED
budget: team quota WITHIN LIMIT
$
8+ yrs regulated infra 4 bank platform deliveries Banking · Fintech · Enterprise SaaS
What I build
The deployment, not the demo
Infrastructure that runs AI in production for organizations where governance and reliability are first-class requirements.
AI & LLM Platform Engineering
Governed platforms for running LLMs and agents in production: AWS Bedrock and LiteLLM gateways, multi-account architecture, access control, observability, and cost governance. The layer that turns an AI experiment into something a regulated org can actually run.
AgentOps & Production AI Reliability
Operating agentic systems safely once they are live: health monitoring, human-gated actions, and audit trails. The discipline of running AI in production rather than prototyping it.
Core DevOps & Cloud Engineering
The full foundation, in any environment: AWS and Azure architecture, Kubernetes, CI/CD, Infrastructure as Code with Terraform and Ansible, and production incident response. Eight-plus years across banking, fintech, and enterprise SaaS.
Production Incidents, Resolved
The hard production failures others have given up on - diagnosed, fixed, and turned into runbooks so the same failure does not recur. Networking, failover, and observability that reflects reality.
8+
Years Experience
4
Bank Platform Deliveries
~30%
Cloud Cost Reduced
15+
Projects Delivered
Who you work with
You work with me directly - no bench, no hand-offs. Currently the lead DevOps engineer delivering a core banking platform into production across four banks, concurrently.
Products
Tools I've built
I build the tools I wish existed for this work - and ship them.
Tendwell
Self-hostable, local-first AgentOps for production health. It observes metrics and runbooks, reasons with a local LLM, and explains what it finds - with human-gated, hash-chained-audited actions. Built for security-conscious and regulated teams.
tendwell — health analysis
local-first · no egress
> is production healthy right now?
availability: error_rate=0.057 BREACHED
propose: restart_service AWAITING HUMAN
audit: proposal recorded HASH-CHAINED
Terraback
Reverse engineer your cloud infrastructure into Terraform code. Terraback imports live AWS, Azure, and GCP resources into clean, maintainable Terraform - so teams can retrofit Infrastructure as Code onto legacy ClickOps environments.
terraback — import
zero-diff plans
$ terraback scan aws --region eu-west-1
discovered: 214 resources across 12 services
$ terraback generate
wrote: modules/vpc, modules/eks, modules/rds HCL
CarVision
A production car-recognition engine: make, model, and generation from a single photo. 896 model-generations across 76 makes at 93.85% top-1 accuracy, with calibrated confidence and explicit rejection handling. Runs on-device (TensorFlow Lite, Core ML, ONNX) in ~50 ms, or as a self-hosted API - it powers Boby's Garage in production.
carvision — inference
on-device · ~50 ms
$ carvision classify photo.jpg
BMW 3 Series (F30) · 2011–2019 TOP-1
confidence: 0.94 CALIBRATED
runtime: tflite-fp16 · 48 ms
Field notes
Latest from the blog
Practitioner write-ups from regulated banking work - running AI in production, compliance in practice, and the infrastructure underneath.
The four-hour clock starts before you understand the incident: notes on DORA Article 17 in practice
Why incident classification is harder than the regulation makes it look, and why the tooling you bring in is itself a compliance surface.
Building AI platforms where reliability and compliance matter?
If you are running - or planning to run - AI in a regulated or high-stakes environment, let's talk about the infrastructure underneath it.
Replies within one business day · NDA-friendly · contact@reops.tech