# AI in Production
Ashwin Bilgi
*NEC Software Solutions — May 2026*
--- About me: An Platforms/Infrastructure engineer learning how to run AI in Production
- 20 years in the industry - Currently Engineering Manager for Product and Billing Usage Data at GitLab - Private Cloud platform at Booking.com - Principal Engineering Manager at Microsoft for Azure Databricks - Senior Site Reliability Engineer at Microsoft for Outlook Mobile - Database Platforms at VMWare - Building https://vlindercli.dev in my spare time. - This talk is from my learnings while building Vlinder. - This talk does not reflect the views of my employer
--- ## What blocks serious AI transformation? ---
---
---
---
**46%** of AI proof-of-concepts scrapped before production.
— S&P Global, 2025
**6%** of companies fully trust AI agents for core processes.
— HBR, 2025
**89%** observe. **32%** still blocked on quality.
— LangChain, 2026
---
--- ## Our tools are built for a different world
Aggregates. Percentiles. Long-tail latencies. Built for the first kind of system.
--- ## Failures compound
An agent with 90% reliability. Six calls in a chain.
**51% success rate.** --- ## The one place it works
Coding agents. - Formal specification. - Deterministic guardrails. - Fast feedback. - An expert human in the loop who owns the outcome.
--- ## But does it, really? Cognitive surrender
Developers want it to work. They retry. They steer. They accept approximations. At some point, acceptance becomes surrender.
--- ## Customers don't write prompts the way developers do.
End user inputs are lossy Especially if they are on mobile.
--- ## Desensitised
Customer support bots. Automated recruitment rejection emails. Your users have been here before. Every bad experience lowered the bar further.
--- ## The real customer
The human whose judgment determines whether an AI interaction succeeded. Not the buyer.
--- # So how do we really run AI in production? --- ## Treat AI as an unreliable external service. What does this mean? --- ## Isolate it
Separate your deterministic code from your non-deterministic code.
--- ## Capture all state
Queues. Distributed logs. Event sourcing. You already build systems where every state transition is recorded. AI is no different.
--- ## A letter
DEL
MUM
HYD
BLR
MAA
--- ## Lost
DEL
MUM
HYD
✗
BLR
MAA
--- ## A copy at every hop
DEL
MUM
HYD
BLR
MAA
--- ## Design failure first
For every AI call: what does rollback look like? Retry? Escalation? Compensating transactions are the real product.
--- ## Build on protocols
HTTP. REST. SQL. You've always built on protocols, not products. Evangelise standardisation.
--- ## Benchmark rigorously
Test plans. Acceptance criteria. Regression suites. You already do this.
--- ## Be able to swap models
You've migrated databases. You've switched cloud providers. Optimize for the ability to swap. Your evaluation infrastructure is your competitive advantage.
---
--- ## The free text box is worth building. It just takes a lot of good old fashioned engineering. --- # Questions? --- # THANK YOU!