// where the hours go

Agents get read access on day one. The engineering is deciding when they get to write.

STACK_01

Agentic AI, in production

A model that answers is easy. A model that acts needs scoped tools, memory it can trust, and a governor that can refuse. Then it has to survive a deploy.

NOW ServiceNow agents: SOP-grounded triage, multi-step remediation via Flow Designer and MID Server. The runbook now proposes its own fix, behind a human gate. NOW RAG on Bedrock: every answer graded against its sources before it ships, so a bad retrieval fails the eval instead of reaching the user. BUILT zero-egress inference: llama.cpp with local RAG on-box, so the assistant runs where the data is not allowed to leave.
STACK_02

The model layer

Retrieval decides whether an answer is true. Chunk, embed, index. Then pull top-k, rerank, and grade the answer against its own sources before it reaches anyone.

RAGAgentic RAGEmbeddingsVector DBsBedrockRAG evalstructured output
ingest docs chunk embed index top-k serve query retrieve rerank grade answer
STACK_03

Cloud, infra & reliability

Written down so it rebuilds without me. If I'm the reason something works, I've built it wrong. Everything idempotent, everything traced.

AWSTerraformKubernetesCI/CDSRELinuxOTEL

            

        

        

            STACK_04
            
Adversarial lens · White Hat

            
I break my own systems on purpose, with permission and a paper trail. Red-teaming something I built is the fastest way to find where the trust boundary actually leaks.

            
ReconWeb ExploitationPrivilege EscalationThreat ModelingPrompt-injection defenseResponsible Disclosure

            

                
                    
                    
                    trust boundary
                    attacker
                    
                    
                    
                    
                    
                    edge
                    app
                    agent
                    data
                    report
                
                

            

        

    


    
SKILLS

    

        

            
            ~/skills
            
            
            
        

        

            

            

        

        
● production◐ shipping○ exploring

    

    

        
GenAI & Agentic RAG, Agentic RAG, Bedrock, AgentCore, Frameworks (Strands / LangGraph), MCP, A2A, Memory, Multi-agent, RAG eval, Observability, Tools, Embeddings, Guardrails & HITL, Context engineering, Governance

        
Cloud & Infra AWS, Kubernetes, Docker, CI/CD, Terraform, SRE, Linux, Ansible

        
Backend Python, Flask / FastAPI, Node.js, REST APIs, Webhooks, ServiceNow

        
Data & ML Ops Classical ML, MLOps, Vector at scale, Eval & observability, Data engineering

        
Frontend React / vanilla, Client crypto

    


    
THE ROAD IN · SEVEN YEARS, THREE NAMES

    

        
2019 to 2023systems
Mostly learning how things break, usually at night, usually on someone else’s schedule.

        
2023 to 2026cloud · devops
Stopped fixing things by hand and started writing down what they were supposed to be instead.

        
2026 onforward deployed
Sent to wherever the problem is. Agents, and the arguments about what they should be allowed to do.

    

    
Seven years and three titles, and honestly the work hasn’t changed much. What changed is how far I sit from the thing that’s going wrong.


    
PAPER THAT AGREES WITH ME

    

        
AWSSOLUTIONS ARCHITECT · ASSOCIATE

        
AZUREDEVOPS ENGINEER

        
CLAUDECERTIFIED ARCHITECT · PROFESSIONAL

        
AGENTIC AIENGINEER · PROFESSIONAL

    


    
LAWS I WROTE FOR MYSELF

    

        
01If it can’t be rebuilt from a file, it doesn’t exist.

        
02The model is the easy part.

        
03Anything irreversible waits for a human.

        
04Write the fix as carefully as the exploit.

        
05Boring, observable, repeatable. In that order.

        
06If I’m the only one who understands it, it’s broken.

    





BACK