Anitha Selvarajan

Problem

After the SVB acquisition, two commercial banking platforms ran side by side, with no agreed path to one.

What I did

Led the comparative analysis of both systems that set the executive migration path, folding Hogan into the Signature platforms.

• Impact
• $65M eliminated over 2 yrs
• Consolidation 6 months faster

• Enterprise architecture
• Executive strategy

Problem

Around 10,000 users across market segments were on an older digital platform and couldn't be moved all at once without disrupting their day.

What I did

Ran a four-year phased migration onto the modern GO platform.

• Impact
• ~10,000 users migrated
• $15.5M/yr program

• Microservices
• REST
• Program delivery

Problem

Payments ran on vendor rails, a recurring cost and a dependency the bank didn't control.

What I did

Replaced them with in-house services on AWS S3, Kafka, and APIGEE, and split the Admin domain into Client, User, and Product microservices.

• Impact
• $6.22M annual savings
• Vendor dependency removed

• AWS S3
• Kafka
• APIGEE
• Microservices

Rebuilt audit-event capture and reporting with Risk, Compliance, and Governance — the unglamorous work that decides how an examination goes.

• 80% better audit readiness
• 50% fewer event gaps

• Governance
• SOC2
• Splunk

Problem

Stock screens hand you a verdict without showing their work, and they never tell you what they could not verify. I wanted a rating I could audit line by line — and six research agents working the same watchlist will contradict each other unless something makes them agree on the rules before any of them starts.

Approach

One agent writes the contract — the nineteen parameters, a 0–10 rubric for each, category weights, decision thresholds, red-flag definitions — and researches nothing itself. Four agents read that single file and run in parallel: technical, fundamental and moat, sentiment, and a separate options agent whose verdict is never blended into the stock score. A sixth joins their findings, weights the four categories 20/25/15/40, and applies a red-flag override built as a one-way ratchet — it can move BUY to HOLD and nothing else. Every value carries a source URL and an as-of date; anything unsourceable is recorded as null rather than guessed.

Result

What you see is today's work, not a screenshot from launch day. Every weekday after the market closes, the agents redo their research and the dashboard redraws itself — ten names, nineteen signals each, every score backed by a dated source you can open. And when I change my mind about what matters, it's a one-line edit, not a rewrite.

Problem

An MCP server hands a model live tools against real data. In a regulated environment “it works” is not the question — what it is permitted to do, and what it leaves behind, is the question.

Approach

An MCP server with a policy layer in front of it: scoped tools, a permission model, redaction of restricted fields, and a per-call audit log. Built around the questions an enterprise architecture review actually asks before granting a permit to develop and operate — because I have sat on the other side of those reviews.

Result

Finished, and it writes its own permit packet. Four principals (a care manager, a claims analyst, an unattended assistant and an auditor) run real MCP sessions against a synthetic health plan. The packet shows what each was allowed, what was redacted or withheld and why, a break-glass access landing in the auditor's review queue, a kill switch taking effect mid-session, and four attempts to tamper with the audit trail, all caught. Every number on it comes from the run; 44 tests keep it that way.