Arun Saravanen
I build AI systems that have to survive enterprise production.
ServiceNow Certified Technical Architect with ten years of Fortune 500 delivery and escalations. I write about agents, retrieval and evals, and about building on ServiceNow.
Ask my notes
A retrieval system over everything I've written. It cites the section it used, and when retrieval is weak it refuses before calling the model. The evals list what it still gets wrong.
35/47evals passingSee the evals
Question
Should a RAG system refuse when retrieval scores are weak?
Sections retrievedgate 3.6
Retrieval has to fail loudly8.33
Introduction6.93
Hold out a query set you didn't build the system for5.81
The anti-pattern I keep removing5.14
Answered with citations
Question
What's your favorite pizza topping?
Nothing I've written matches, so Ask stops here. No model call, no guess.
Refused before any model call
Built and running
Public evals
Hand-checked cases run against the exact pipeline visitors use. Failures stay on the board.
- Refusals40/47
- Retrieval24/31
Draw a digit
A small neural network that runs in your browser. Draw a number and it guesses which one.
What I bring to a team
AI measured in public
Retrieval and agent systems with guardrails, scored by evals anyone can read, failures included.
35/47 evals passingTry Ask
Enterprise delivery under pressure
Ten years with Fortune 500 teams, from executive escalations to platform architecture.
30+ Fortune 500 engagementsSee my path
Tradeoffs, explained plainly
I translate technical depth for engineers and executives alike, and I write it down.
8 posts on AI and ServiceNowRead the writing
Writing
View all- Neural Networks 101: How a Computer Learns to Read Your HandwritingAI
A simple, visual walkthrough of how neural networks recognize handwritten digits, inspired by 3Blue1Brown's classic Deep Learning series.
- RAG That Doesn't Lie: Grounding Patterns I TrustAI
Most RAG systems hallucinate when the retriever fails silently. Here are the four grounding patterns I lean on in production and the one anti-pattern I keep removing.
- Flow Designer at Scale: Patterns That Survive RefactorsServiceNow
Six patterns I keep reaching for when a Flow has to outlive the team that built it: naming, idempotency, error handling, and the parts I always extract into subflows.
- Building Agent Loops That Don't SprawlAI
A short, opinionated checklist for keeping an agent loop bounded: tool budget, termination conditions, and the single observability hook that's saved me the most debugging time.
- CMDB Health: An Honest Audit PlaybookServiceNow
Dashboards lie. Here's the audit I actually run when a stakeholder asks 'is our CMDB any good?' Five queries, three interviews, and a follow-up I always regret skipping.
Let's talk
If you're building AI that has to hold up with real customers, I'd like to hear about it.