Systems & Prototypes

Projects

Systems, tools, and experiments from the frontier of AI engineering.

10 systems

Open Source

My own work, and public
Open Sourcecontext-grid
Personal Project · 2026

A test bench for the machinery that feeds documents to a language model. Almost every rule of thumb in this field is advice nobody measured, so this measures it on your own documents and your own questions, and ranks what comes out on quality, speed and cost. The answer key is stored as a stretch of characters in the original document rather than as a chunk number, so a set of test questions written once stays correct even after you change how the documents are cut up.

On PyPI under MIT, still an early release — it turns retrieval folklore into numbers on your own documents.

Read the docs →
PythonRAGRetrievalEvaluationBenchmarkingCLI

Enterprise

Built at work, so the details stay inside
Enterprise / ConfidentialAgentic Ticket Resolution Recommender
ConnectWise · 2024 – 2025

An agentic AI system built on the Agno framework that analyzes millions of historical support tickets and recommends resolutions in real time.

60% reduction in resolution time, 40% drop in support costs

AgnoAgentic AIRAGPythonFastAPI
Enterprise / ConfidentialConfig-Driven Agentic AI Framework
ConnectWise · 2024 – 2025

A reusable framework for deploying config-driven AI agents via REST API, with support for memory, multi knowledge base retrieval, tool use, and enterprise integrations.

Adopted across multiple internal products and teams

Agentic AIMCPLLMsFastAPIMLOpsRAG
Enterprise / ConfidentialIn-House RAG Framework
ConnectWise · 2024 – 2025

A low-code RAG framework that lets teams deploy retrieval augmented apps quickly, adopted organization wide by engineering and product.

Adopted by all internal teams and product managers

RAGLLMsVector SearchPython
Enterprise / ConfidentialDomain-Pretrained Language Model
Draup · 2021 – 2024

A RoBERTa based language model pre-trained from scratch on 60M job descriptions to lift downstream NLP performance in the labor market domain.

5%+ gain on downstream NLP tasks

TransformersPyTorchNLPRoBERTa

Experiments

Built out of curiosity, not for work