An enterprise engineer who moved into AI, not away from engineering
I spent fourteen years inside enterprise software products where a bad release reached real customers. That is where my opinion of AI systems comes from.
I joined CA Technologies in July 2008 and worked across CA APM, Unicenter NSM, Workload Control Center and AutoSys — quality engineering, automation, release, security and the kind of customer escalations where you get a crash dump and a deadline. I led regression automation, contributed to a major 32-bit to 64-bit product conversion, and fixed SQL injection, command injection and cross-site scripting defects that had been sitting in the code for years.
At Progress Software from 2017 I moved closer to the platform: infrastructure as code, installers, CI/CD, and continuous validation for enterprise products. That work produced the numbers I still quote — a validation cycle cut from roughly 72 hours to 6, a recurring permissions operation from about 9 hours to 5 minutes, regression effort per release from 5760 hours to 640.
Since August 2022 I have worked independently, and since founding Sai Ameya Technologies LLP I have operated under the Ameya Labs name. My work moved to AI in the only way that made sense to me: as an architecture problem. Generative models are probabilistic components with unusual failure modes. Everything I know about contracts, evidence, observability and release discipline applies to them — and most of it is missing from the way teams adopt them.
Six principles I actually apply
These are not values on a wall. Each one has cost me a project or saved one.
Define the business problem before choosing technology
If I cannot state the decision this system changes, and who makes it, I am not ready to design it. Model choice comes last, not first.
Make scope and trade-offs explicit
Every system is bounded by something: data quality, latency, cost, permissions, regulation. Writing the boundary down turns a surprise into a decision.
Keep facts traceable to sources
If a number or a claim appears in a product, someone must be able to walk back to where it came from. This applies to my own site as much as to client systems.
Use deterministic computation where correctness requires it
Totals, eligibility, retention periods and permissions are calculated, not inferred. A model may explain a result; it does not get to invent one.
Validate implementations with meaningful evidence
A passing test suite proves the tests passed. Real evidence is a build, a diff, a measured figure, a screenshot of the actual state, or a reviewer who re-checked the source.
Preserve human control over consequential decisions
Autonomy is cheap to grant and expensive to withdraw. Deployment, disclosure, money and personal data stay behind an explicit human approval.
How I prefer to work
Whether the engagement is a review, an architecture, or a build.
Start with the problem and the constraints
A short written problem statement, the constraints that are not negotiable, and what 'done' means. This is usually the most valuable artifact of the whole engagement.
Design the boundaries before the features
What the system may decide, what it must never decide, where data is allowed to go, and what happens when a component fails.
Ship with the evidence attached
Handover includes the reasoning, the trade-offs, the checks that were run and the ones that were not. You should be able to maintain it without me.
Education
- 2006 – 2008
- Master of Science, Information Technology — IIIT Hyderabad
- 2002 – 2006
- Bachelor's degree, Computer Science & Engineering — JNTU Hyderabad