Mirrell M. Knabe III

Projects

AI applications I've built, and the enterprise programs I've architected, from cloud platforms to mainframe modernization.

Animation: an AI robot fading into a cloud, then into a mainframe

Enterprise programs

Selected engagements from my career with Cognizant, BNY Mellon, Highmark, Accenture and PNC Bank.

AI · Cloud

Credit Card Portfolio Integration & Migration

Architect · Cognizant · 2026 – present

Fortune 500 financial services leader. Leading Claude AI integration into mainstream development, designing agentic workflows for credit decisions, disputes, fraud and refunds, and directing AWS conversion and data loads for the acquired card portfolio.

Cloud · AI-assisted

Retail Mainframe Zero Program

Migration lead · Cognizant · 2025

Fortune 500 retail leader. Migrated 40+ applications from mainframe to Python, Java, AlloyDB and BigQuery on GCP, with AI-assisted development speeding delivery 40–60% and JCL batch converted to Python.

Distributed · Event-driven

Corporate Trust & Trades Modernization

Integration architect · Cognizant · 2019 – 2024

Fortune 500 banking leader. Kafka event-driven architecture for real-time trade status, payments and messaging; MQ-to-Kafka pipelines; generative AI for Java NLP APIs and automated code review.

Distributed · Data

BNY Mellon Treasury & Fixed-Income Modernization

Sr. developer & analyst · BNY Mellon · 2014 – 2019

Moved Treasury and Fixed-Income systems from mainframe to SQL Server with .NET/C#, ETL pipelines and regulatory reporting for CCAR and Dodd-Frank.

Healthcare · Modernization

Highmark Claims Processing Modernization

Consultant & sr. developer · Accenture / Highmark · 2006 – 2014

Converted claims, member and provider systems to web-enabled Java/AngularJS applications and modernized 3,500+ COBOL batch and CICS programs, with HIPAA-compliant data integration.

Mainframe · Compliance

PNC Patriot Act System

Team lead · PNC Bank

Led an 11-engineer team that delivered a mainframe system for identifying suspicious monetary transactions within a 90-day regulatory deadline.

AI applications & tools

FlowForge desktop app showing a 12-node card dispute triage workflow with Claude AI agents
Shown: a 12-node card-dispute triage workflow. Click to enlarge.
Featured project

FlowForge — Agentic AI Workflow Designer

A desktop tool for visually designing and prototyping AI automation workflows. Drag and drop 25 node types (Claude AI agents, triggers, integrations, logic and reports), connect them into pipelines, and run them end to end, in simulation or with live Claude API calls for the AI steps.

  • Branching logic, parallel steps and workflow variables
  • Live execution log with step-by-step status
  • Four starter templates; save and import workflows as JSON
Python
Tkinter
Claude API
Agentic workflows
Featured project

Model Inspector Studio — Local AI Model Explorer

A desktop tool for looking inside open-source AI models without loading them into memory. It reads safetensors, GGUF, ONNX and PyTorch files, plus full Hugging Face model folders, and shows the configuration, the layer architecture, every tensor and a visual map of where the parameters live.

  • Model Map: parameter breakdown by layer group, from its own parser
  • Searchable tensor table with shapes, data types and parameter counts
  • Safe by design: never unpickles PyTorch files automatically
  • Scans drives to find models; one click opens the full graph in Netron
Python
Tkinter
safetensors · GGUF · ONNX
Netron
Model Inspector Studio showing the parameter map of DeepSeek-R1-0528-Qwen3-8B
Shown: Model Map of DeepSeek-R1-0528-Qwen3-8B (GGUF). Click to enlarge.
Local LLM Test Bench showing DeepSeek-R1-0528-Qwen3-8B's reasoning trace for a logic puzzle
Shown: DeepSeek-R1-0528-Qwen3-8B reasoning through a logic puzzle, running locally. Click to enlarge.
Featured project

Local LLM Test Bench — Private Model Testing

A desktop chat app for running and evaluating open-source AI models entirely on your own machine through llama.cpp. Responses stream in token by token, and for reasoning models such as DeepSeek-R1 the chain of thought and the final answer appear in separate, clearly marked sections.

  • Fully local: no prompts or data leave the machine, suited to sensitive work
  • Tunable context size, temperature, top-p, max tokens and GPU offload
  • Reasoning trace can be shown or hidden with one click
  • Stop generation, start a new chat, or copy the last answer
Python
Tkinter
llama.cpp
Local LLMs (GGUF)
AI platform

AI Innovation Hub

An AI application suite, re-architected from a Dockerized FastAPI and React web app into a Python desktop suite.

Modernization tooling

Mainframe Modernization Studio

A multi-stage pipeline for converting COBOL, JCL, IMS and DB2 assets, wrapped in a FastAPI service.

Web tooling

Site Builder Studio

The drag-and-drop builder used to create this site. It exports plain HTML and CSS for static hosting.