Hyderabad, India · Enterprise AI platforms
I'm Soumendra Kumar Sahoo. Over fourteen years I've moved from mainframes to Big Data, multi-cloud and AIOps at IBM, enterprise search, MLOps and LLMOps. Today I own the layer that decides whether an AI system survives real users: architecture, instrumentation, evaluation, governance, and the people who run it. I still read the traces myself.
Fourteen years · six organisations
Since 2024
2021 – 24
2018 – 21
2016 – 18
2014 – 16
2012 – 14The operating layer
Most AI failures are operating failures. What breaks in production is evaluation nobody owns, instrumentation added the week after the incident, and governance written by people who have never read a trace. This is the span I work across.
01
Start from the decision the organisation needs to make, the risk it needs to control and the feedback it needs to learn. Architecture follows those constraints. The goal is a shared capability several teams can use, instead of another heroic one-off every quarter.
02
Traces, cost attribution and a shared failure taxonomy, agreed before launch rather than negotiated after an incident. Instrumentation judgment decays fast when you stop doing it yourself, so I run self-hosted Langfuse on my own infrastructure and keep the certification current.
03
Evaluation is a standing mechanism with an owner and a cadence, running long after launch. It also needs evaluating itself. LLM-as-judge carries real bias and a real bill, and retrieval architecture changes what "correct" even means.
04
Policy, accountability, monitoring and continuous improvement assembled into one operating system rather than a document nobody opens. Autonomy earns bounded permissions, verified backups and tested recovery, the same controls any other production system gets.
05
Mentoring, written decisions, interview design that surfaces real understanding, and influence that works without a reporting line. Open source taught me the last one. You can't assign a task to a volunteer community, so the mission has to carry itself.
The record
Each one is verifiable against my profile, my publications, or the archive on this site.
14+
years in enterprise engineering, 2012 to today
6
organisations, from TCS to PepsiCo
115
posts published since May 2012
OTCA
OpenTelemetry Certified Associate
Selected writing
I write to make architecture choices, leadership patterns and failure modes easier to inspect.
What failed
Most leadership sites publish only outcomes. A standing failure column is harder to fake, and it shows how someone reasons when the system doesn't cooperate.
Eleven invented model names and four fake paper titles, produced with total confidence. The fix was a designed verification control. Better prompting would never have caught it.
Week 20, 2026No alert, no verification step, nothing to restore from. Agents need bounded permissions and tested recovery like any production system.
Week 31, 2026First attempt, strict remote proctoring. The second attempt worked because I changed the preparation method rather than adding effort.
January 2026A few hours with AI tools produced a working prototype for a problem I didn't actually have. Building isn't the same as solving.
January 2026Let's compare notes
Leadership roles, advisory, speaking, community work, or a straight argument about evaluation design. Send a note and I'll reply by email.
New here? How to work with me covers what to put in a first message.
Prefer your own client? contact@soumendrak.com