Nripesh Kumar — Product Manager

Products I've built — from problem to shipped solution

Each case study shows the user problem, my product approach, and a measurable outcome. The first is a product I shipped as an engineer at Visa; the rest are self-directed AI products built solo for this portfolio.

V-ARM · Visa's B2B virtual card payments platform

$17.8M

in virtual card transaction volume, V-ARM's first month

50%

less localization effort

3 mo

to full capability

Shipped @ Visa

V-ARM — B2B Virtual Card Payments

Commercial Money Movement Solutions · translated product requirements into the integration contracts and APIs that launched V-ARM.

Problem

Launching a B2B virtual card platform meant turning ambiguous product requirements into integration contracts and APIs that issuers, processors, and engineering could build against — while two separate schedule risks threatened the launch date.

Solution

Flagged a 20% throughput loss early enough to ship a phased MVP on time, then recovered a one-month security delay by restructuring sequential work into parallel execution across Engineering and Cybersecurity. Later rebuilt localization as a configuration-driven framework to unlock the European launch.

Outcome

$17.8M in first-month virtual card volume · on-time launch despite two schedule risks · new markets with zero code changes

$17.8M

FIRST-MONTH VOLUME

50%

LESS LOCALIZATION EFFORT

3 mo

TO FULL CAPABILITY

PM insight: Owned the requirements-to-API translation layer end-to-end, and protected the launch date twice — by catching risk early, and by renegotiating how the team would deliver, not whether.

Enterprise B2B Fintech API Design Cross-functional Leadership

Self-built · Shipped

PRD Quality Reviewer

Scores a pasted PRD against the SMART framework with section-level reasoning and suggestions.

Problem

Product managers spent 2 hours reviewing each requirements document for clarity, completeness, and edge cases.

Solution

Built an AI pipeline using the Gemini API to score PRDs across 5 SMART dimensions with per-criterion feedback and improvement suggestions.

Outcome

2 hours → 2 minutes · 60x faster · clear, actionable per-section feedback

60x

FASTER REVIEW

SMART DIMENSIONS

0→100

SCORED OUTPUT

PM insight: Built after designing the same pattern at Visa — 6 parallel AI agents that cut BRD review from 2 hours to 2 minutes. This is the public version: Gemini-based SMART scoring anyone can try.

Next.js Gemini API TypeScript Tailwind CSS

Self-built · Shipped

Payment Instruction Parser

Extracts structured payment data from messy .txt and .eml emails with confidence scores.

Problem

Finance teams receive payment instructions through messy email threads, forwarded invoices, and informal reminders — manual extraction risks missing due dates, vendors, or payment methods.

Solution

Built a finance tool that accepts email uploads, extracts structured payment instructions via LLM structured outputs, and stores results in PostgreSQL for review and audit.

Outcome

Reduces manual data entry · handles multiple payments per email · audit-ready extraction history

100%

AUTOMATED EXTRACTION

Multi

PAYMENTS PER EMAIL

Audit

READY HISTORY

PM insight: Rooted in regex-based email parsing I shipped at Visa — I logged the tradeoffs of swapping brittle rules for an LLM and where hallucination risk still requires a human check.

Python FastAPI Streamlit PostgreSQL

Self-built · Shipped

B2B Payments Dispute Classifier

Triage assistant for bank dispute analysts — AI that recommends, never decides alone.

Problem

Dispute analysts spend 15–30 minutes per case reading unstructured dispute text and cross-referencing resolution playbooks — roughly $125K/month in analyst time for a mid-size bank, with inconsistent outcomes across analysts.

Solution

Built an AI triage assistant that classifies disputes, suggests a resolution path, and flags regulatory requirements in seconds — using confidence-tiered routing (HIGH/MEDIUM/LOW) so low-confidence or contradictory cases are automatically blocked from automation and routed to mandatory human review.

Outcome

15–30 min → seconds per case · AI as assistant, not decision-maker · full audit trail for compliance

3-tier

CONFIDENCE ROUTING

$125K

MONTHLY COST AT STAKE

100%

HUMAN REVIEW ON LOW CONFIDENCE

PM insight: Shows how to handle model uncertainty in a regulated industry — confidence thresholds, mandatory escalation, and auditability, not just a clever prompt.

Next.js Gemini API TypeScript Tailwind CSS

Experience

Jul 2022 → Jul 2026

Visa

Senior Software Engineer, Commercial Money Movement Solutions

Atlanta, GA

Launched V-ARM, a B2B virtual card payments platform, translating product requirements into integration contracts and APIs, enabling $17.8M in first-month virtual card transaction volume.
Protected V-ARM's launch date by flagging a 20% throughput loss and proposing a phased MVP, launching on time and reaching full capability within 3 months.
Recovered a one-month V-ARM security delay by restructuring sequential work into parallel execution across Engineering and Cybersecurity, preserving full scope and the original MVP launch date.
Unlocked V-ARM's European launch by rebuilding localization as a configuration-driven framework, cutting localization effort 50% and enabling new markets with zero code changes.
Cut BRD review time from 2 hours to 2 minutes for Product and Engineering by designing 6 parallel AI agents that score requirements and recommend rewrites.
Improved engineering velocity by launching a repository-aware GenAI assistant adopted by 60+ developers, increasing deployment frequency 25% and reducing time to pass unit-test coverage and security checks 30%.
Mentored 4 engineers in Visa Startup Bootcamp through product ideation, AI/payment-flow prototyping, and pitching, helping the team place 3rd among 20 teams globally.

Jun 2021 → May 2022

Columbia University

Graduate Research Assistant, E3B Dept. — Environmental Science & Policy

New York, NY

Accelerated satellite imagery delivery by designing a configurable ingestion pipeline around researcher requirements and API constraints, processing 2TB while increasing daily ETL throughput by 30%.
Automated manual GIS workflows to produce analysis-ready maps from overlapping satellite imagery, reducing researcher hours by 20% and freeing time for deforestation analysis.

Oct 2020 → Dec 2020

Flatworld Solutions

Analyst Intern, global automation & AI consulting firm

Bangalore, India

Reduced mortgage application processing time by 40% by automating data extraction with Amazon Textract and designing custom table parsing with human review for ambiguous fields.

Education

2026 → Jun 2028

UCLA Anderson School of Management

M.B.A., Full-Time Program

Los Angeles, CA

705 GMAT Focus (750 GMAT equivalent). Merit-based Fellowship Awardee.
Concentrating on product management, technology strategy, and entrepreneurship.
Active member: AnderTech, Management Consulting Association.

2021 → Jun 2022

Columbia University

M.S. Computer Science, ML Specialization

New York, NY

Machine Learning specialization with focus on NLP and computer vision applications.

2016 → Jun 2020

National Institute of Technology, Trichy

B.Tech. Honors, Computer Science

Trichy, India

Economics Minor.
Organized Ethical Hacking workshop and Blind Coding competition (Pragyan Tech Fest).
Led technology awareness workshops for school students (National Service Scheme).

Built the infrastructure.
Now I want the roadmap.

Engineer who thinks in products. Builder who ships.