Santhosh R
AI Product Manager · Senior Business Analyst — Banking · Bengaluru
I build the AI systems I specify.
Eight years as a techno-functional business analyst across banking, healthcare and enterprise technology, with an MSc in Data Science — and, unusually for a BA, I ship the thing I wrote the requirements for.
The work below has one thing in common: the interesting decision was never which model to use. It was deciding what the model was allowed to do, who signs off before output reaches a person, and what the whole thing costs per call.
Case Studies
Three products. Each opens the full walkthrough in a new tab.
AI Trading Copilot
A market-analysis product where the language model is never allowed to state a number.
Efrtly
Post-ready captions, hooks and hashtags — grounded in what performs, not what a model imagines.
Regulatory Reporting
Fixed price, no agreed process, clock already running. So I built the process before the features.
Find Where It Breaks
Migrating a hospital group's reporting suite, and tracing every mismatch to the layer that caused it.
Evaluation and guardrails, in that order.
Two questions decide whether an AI feature is ready: how do you know it works, and what stops it when it doesn't. Most demos answer neither. Each case study above answers both, including where my own answer was inadequate.
Draw the boundary at the numbers
In the trading product the model cannot emit a figure — it writes prose around numbers computed elsewhere. A hallucinated price and a trade are separated by architecture, not by a prompt.
A person signs off on the change
Nightly calibration can propose new scoring weights. It cannot apply them. On Efrtly, nothing published without the creator approving it. Automation proposes; a human decides.
Publish the number, sample size included
45.7% directional accuracy at T+3 across 6,921 observations, on the page where users can see it. A hit rate without an n beside it is marketing.
Audit the measurement itself
I had my own scoring and calibration code reviewed. It found the information coefficient was pooled across dates, so I void every historical value rather than quote it.
Name the gap before someone finds it
No measured groundedness rate on generated claims. No adversarial testing. Outcome logging built fifth instead of first. All three are written into the case studies.
Treat spend as a product metric
Thirteen AI workloads, each writing a per-call cost log that reconciles to the provider's billing. $6.40 to $1.60 a day across all of them, with no change a user would notice.
Publications & Links
Things you can open right now without asking me for anything.
Work Experience
Eight years across banking, healthcare and enterprise technology.
Skills
What I'm actually used for.
AI Product & Evaluation
Enterprise Delivery & Governance
Technical
Education
Let's talk.
Open to AI Product Manager and Senior / Lead Business Analyst roles — on-site, hybrid or remote in Bengaluru, available immediately.