Work

Ratchetloop

Auditable software delivery for AI coding agents. A coding agent works in an isolated Git worktree, deterministic checks validate the change, a different model family reviews it, and a person makes the merge decision. On one day, fifteen work items went through it from build to merge, including changes to its own code.

Case study · Public repository · The method behind it

langgraph-factory

Multi-agent code generation on local models: a planning model owns the architecture and a coding model generates and repairs, with a recovery ladder that tries cheap mechanical fixes before spending model calls. No API keys or per-token cost. A sample run produced a working 24-file blog platform in 266 seconds.

Python Algorithm Interview Work Sample

Ten representative algorithm patterns built through Ratchetloop and validated with focused unit tests, deterministic reference-oracle tests, cross-model review, and GitHub Actions.

AI coding agents in daily practice

Claude Code, Codex CLI, Cursor, GitHub Copilot, and other agents used daily and matched to the task, plus MCP servers and CLI tools built for agent workflows. Hands-on onboarding of other engineers to agentic development, and AI-assisted development inside a current enterprise client's official GitHub Copilot pilot, under its review and compliance rules.

Verizon, 2012 to 2024

Senior consultant and lead engineer delivering customer-experience systems for Fortune 100 and public-sector clients, working directly with client teams, executives, and security reviewers.

• Designed SolVE, a project-estimating and cost-validation tool tied to Salesforce opportunities. It was adopted across Professional Services and later handed to internal IT.
• Led delivery of a web application for the New Jersey Department of Transportation that brought in six-figure new revenue, plus API services connecting web, IVR, and chatbot clients to back-end systems and SAP.
• Led in-house development of mobile apps and an SDK (iOS, Android, WebRTC) for a new Verizon product, saving six figures in outside costs.
• Led a large team building an enterprise VoiceXML IVR with Java middleware for a healthcare client, and a small team delivering a custom billing system for a hosted call-center client.
• Ran discovery directly with clients, pushed back when the request did not match the need, and explained tradeoffs to engineering and business decision-makers. Verizon Ovation Award, 2014.

Kaiser Permanente, state agency, and Walmart work is under Contact center & conversational AI.

Current enterprise contract

Production Azure Functions in C#/.NET for a large enterprise client, shipped through Azure DevOps pipelines and the client's change control, working against Oracle and SQL Server.

Intellivoice: email by phone

Led the product end to end: hired the team, built the development setup, chose the hardware and software, and directed delivery through on-site installation. It reached 50,000+ daily users, and its success led Premiere Technologies to acquire the company.

Speech-technology leadership

Director-level roles at speech-technology companies: launched a speech solutions group, led application development teams, built a reusable VoiceXML application framework, and led sales engineering for speech-to-text products.

Kaiser Permanente IVR framework

Technical lead on a configurable VoiceXML call-handling framework for roughly 100,000 calls per hour, with regional customization, CTI screen pops, speech self-service, and high availability under HIPAA constraints. The initial delivery led to multiyear contract extensions.

Conversational AI for state agencies and healthcare

Directed evaluation and integration of Google Dialogflow CX and ES with NICE CXone, NICE inContact, and Genesys: intent design, intent coverage generated from source documents, ASR tuning, bot-to-agent handoff, automated intent testing, and delivery through public-sector security review. A FAQ-bot proof of concept helped retain a $5M annual contract.

Walmart customer support IVR

Designed and deployed the IVR for Walmart's primary customer-support number on NICE inContact, and moved it to the new platform on site without interrupting service.

NICE CXone and conversational-AI bots

Inbound and outbound IVR scripting on NICE CXone Studio, connected to Omilia conversational-AI bots for self-service and bot-to-agent handoff, for a current enterprise client.

Production VoiceXML applications

• The Home Depot, employee time reporting: designed, built, and supported the production VoiceXML application overnight employees used to report working time. It ran for about two years and was the main source of revenue for Speech1st Solutions during that time.
• Comcast field installation reporting (for Voxeo): a VoiceXML application that let Comcast field technicians report completed installations.
• Constellation Pipelines, test-taker authentication: a VoiceXML telephone application that verified test takers' identity.

Speech platforms and standards

Led VoiceXML 2.0, 2.1, and SRGS conformance testing for the VoiceXML Forum across five voice platforms, including Genesys and Nuance. Built speech-to-text dictation solutions and the measurement methods behind them, the work that produced five granted patents. Principal developer of the Davey Tree voice field-reporting application, a Speech Technology Magazine Most Innovative Solution.

Patents and award references

Strategy Assayer

A large-scale computational research and evaluation platform with parallel experiment execution, directed search, walk-forward testing, control arms, provenance, and blocking promotion gates. Systematic trading is the current workload; the underlying architecture is a general research and evaluation system.

Case study

lean-optimizer-public

The optimization core of Strategy Assayer, published to read and run on its own. A 679,140-point parameter search estimated at 88 hours as an exhaustive grid finished in 2 minutes 40 seconds, head to head on the same strategy, ranges, and data. Two gains compound: compile-once caching with persistent .NET workers (about 1.8x throughput) and an incremental genetic algorithm.

GPU forecasting platform

Data ingestion, feature engineering, training, and backtesting with an XGBoost and LSTM ensemble, with GPU memory managed through TensorFlow and CuPy.

Early applied AI: fuzzy logic, neural networks, genetic algorithms

At Togai InfraLogic, an early applied-AI company: a fuzzy-logic airbag controller, built with Autoliv, that beat BMW's production controller in simulation; a genetic algorithm that tunes fuzzy controllers automatically; and about a year with Siemens' fuzzy-logic research group in Munich. Published through IEEE, NASA, and the U.S. Department of Energy.

Asset Factory

Generative-AI production and publishing platform spanning persistent model serving, multimodal generation, quality scoring, evolutionary optimization, human review, packaging, metadata, and marketplace publishing. Assets Forge is one brand produced through the system.

Case study with UI and generated-output gallery

On-premises inference platform

A local model-serving stack on an NVIDIA DGX Spark for cost control and data custody: several open models served side by side, shared GPU memory management, multi-model worker pools, and regression-gated model promotion. Every GPU project on this page runs on it.

Field note: putting a local model to work

Clinical-note workflow trial

Speech transcription, AI formatting of medical notes, and an open-source electronic health record platform in one workflow. Stored-record accuracy improved 57% in the trial environment.

Serverless RAG on AWS

A personal project alongside the AWS certifications: FastAPI, React, Amazon Bedrock, Lambda, and API Gateway with AWS SAM, plus embedding and clustering workflows for grouping similar inputs.

Meetup Finder

A group enters where everyone is, and the app suggests a fair place to meet. No accounts, no app install, no location tracking, and no paid map services.

Case study · Live beta

Route Forge

Multi-tenant route-planning SaaS: efficient routes for 10+ stops with adjustable time at each stop, live traffic, and arrival estimates. Built as three apps (planner, storefront, and API with sign-in) sharing one login, on Next.js, TypeScript, Supabase, and Stripe.

Groundwork Sites

A small-business website pipeline from lead to live site: lead discovery, site generation from templates, Stripe payment, automatic domain registration and DNS, deployment, and a customer self-service portal with role-based access.

Real Estate Market Intelligence

A private web application that a working real estate agent uses every week. It loads regular MLS exports, now more than 100,000 listings across north Georgia, into one local database and turns them into the agent's working material: expired-listing prospect letters with printed envelopes, comparative market analyses, area trend reports, farm-area sheets, and buyer briefs. The weekly prospect mailing goes from a fresh MLS export to addressed letters and envelopes in one step, with a sample sent straight to the office printer for review.

Private application; MLS and client data are not publicly exposed.

Earlier business systems

A data-entry system whose rollout drove a leasing business's growth and profitable sale, a point-of-sale and inventory system for a cellular retailer, and mobile and voice apps for consumer devices.