Skip to content
Vijay Kumar

Work

From promise to proof

Agentic systems first, each with the engineering proof behind its claims. Then delivery case studies, then a compact archive of earlier product, data, and audit work.

Showing all work

Agents & Applied AI

Agentic systems

Each card leads with the problem and the engineering proof behind it. What is simulated is disclosed under Scope.

Agentic AI · Service operations

Refund Assurance — the system that reopens a case when “refund processed” was a lie

Problem: Async refunds fail silently at the gateway; the session still counts as resolved, and nobody knows until the customer calls back.

  • 73/73tests passing
  • 1idempotent write
  • 8step verified pipeline
  • TypeScript
  • React
  • Zod
  • Vitest
  • Playwright

Agentic AI · Accounts payable operations

MatchPoint — resolving three-way-match exceptions inside CFO authority, and stopping exactly where humans must decide

Problem: Price variances, partial receipts, and duplicate invoices stall AP posting; the judgment calls live in policy, not in anyone's head.

  • 201/201tests passing
  • 22Playwright e2e
  • 5additive SQLite migrations
  • 1hash-chained, tamper-evident ERP audit log
  • TypeScript
  • React
  • Express
  • SQLite
  • Playwright
  • Coolify-packaged container

Agentic AI · Security operations

Business Impact Gate — the pre-flight check before an analyst isolates a production SAP gateway

Problem: A containment action can be technically right and operationally harmful; the context that would change the decision lives in other systems, is sometimes stale, and is joined by hand under pressure.

  • 92/92tests passing
  • 8/8browser e2e
  • 0endpoint mutations
  • 2independent stateful simulators
  • TypeScript
  • React
  • Express
  • SQLite (three owned databases)
  • Playwright

Agentic AI · Operations (marketplace)

MarginPulse — pausing paid demand when the kitchen is drowning, and proving the pause stuck

Problem: During a rush, the kitchen and the ad budget live in different systems; pausing means noticing, switching tools, and later checking it actually stopped.

  • 81 + 6Vitest + Playwright tests
  • 1write path
  • 3failure classifications handled without blind retries
  • TypeScript
  • React
  • Fastify
  • SQLite
  • Playwright

Applied AI · Retrieval-augmented generation

GenAI agentic RAG system

Answering questions over a document set quickly and accurately.

  • <2sresponse time
  • 90%+accuracy
  • LLaMA 3
  • RAG
  • ReAct
  • Streamlit

Case studies

Delivery case studies

Open a row for my role and the engagement details.

01Configit

98%

Engagement details

My role
Solutions Consultant → Senior Solutions Consultant, owning end-to-end solution delivery.
When
2021 – 2024
Capabilities
Enterprise implementation · Integration design · Customer discovery
Stack
Salesforce · SAP ERP · Configit CPQ · REST · SOAP · GraphQL · Microsoft Azure · BPMN

02Third Estate Ventures

49K

Engagement details

My role
Business Intelligence & UX Designer, from analyst discovery through implementation.
When
2025 – Present
Capabilities
User research · Data design · Analytics
Stack
Python · REST APIs · Snowflake · SQL · Power BI · Power Automate

03AI Contract Analyzer

70%

Engagement details

My role
Designed and built the system.
Capabilities
Applied AI · Technical problem solving
Stack
Mistral 7B · Ollama · PyMuPDF · Python

04University at Buffalo

Identified user segments and experience drop-offs, and turned them into prioritized onboarding and UX recommendations.

Engagement details

My role
Product Strategy & UX Intern, research through recommendation.
When
2025
Capabilities
Product discovery · Customer research · Prioritization
Stack
Google Analytics · Figma

Archive

Earlier projects

Coursework, prototypes, and case analyses, kept for reference.

ProjectDescriptionOutcome
Nexus event management platform
  • Product discovery
  • Event analytics
Centralized bookings, payments, and coordination; later extended with mobile-responsive design, Stripe and PayPal payments, and analytics dashboards.Centralized event operations
Vespera
  • Product concept
  • UX design
Product and UX concept for a premium automated home cocktail machine.Product concept & UX
LiveSafe
  • Product concept
  • Data visualization
App concept combining real-time crime statistics with quality-of-life metrics.UX design

ProjectDescriptionOutcome
Retail analytics dashboard
  • ETL
  • Power BI
Data warehouse with ETL workflows feeding sales dashboards.Warehouse + dashboards
Word problem solver
  • SVM
  • Optimization
Solves algebra word problems with SVM classification and IBM CPLEX optimization.35% faster solution time
Housing value prediction
  • Regression
  • Prediction
Multiple linear regression on 5,802 West Roxbury housing records.R² 0.79 · test RMSE 46.39
Personal loan acceptance classifier
  • Classification
  • Machine learning
k-NN and logistic regression on the Universal Bank dataset.F1 0.95 (k-NN) · 0.92 (logistic)
Vehicle price prediction
  • Decision tree
  • Prediction
Decision-tree model for Toyota Corolla auction pricing.RMSE 1,302
Airfare price prediction
  • Forecasting
  • Python
Price-forecasting models in Python and pandas on 638 airfare records.Forecasting model
Hostel management system
  • Web application
  • MySQL
Hostel administration with secure registration and notifications (HTML, CSS, JSP, PHP, MySQL).Web application

ProjectDescriptionOutcome
Smithsonian Banking Group
  • COBIT 5
  • Risk assessment
COBIT 5–aligned risk assessment of an internet banking system, including business rules and SIEM recommendations.Control recommendations
ClaimProof Insurance
  • IT assurance
  • Change control
IT assurance audit of change control using COBIT 5: OTP storage, log reconciliation, change management.Control recommendations
Major Manufacturing
  • COBIT 5
  • ERP security
COBIT 5–aligned governance and ERP-security risk assessment.Restructuring & security recommendations
Upward Bound Airlines
  • Cloud migration
  • COBIT 5
Risk assessment of a cloud migration using COBIT 5.Cloud security recommendations
AI for cybersecurity defensive controls
  • Machine learning
  • Threat detection
Applying machine learning, reinforcement learning, and NLP to threat detection and response.Research case study