How I build
AI engineering, in practice.
The useful engineering decisions sit between the model and the person using it. These are a few from products I've built.
Give the assistant a specific job
At KiNDD, a 501(c)(3) nonprofit I co-founded, families need help navigating developmental disability services. Their location, Regional Center, and available providers shape what an answer should say.
I built the web, iOS, and Android applications around a shared Python/Django backend. PostGIS handles geographic matching. Bedrock agents draw on that product context to help families find a next step, with English and Spanish support.
A separate research retrieval path uses Cohere and Pinecone to return cited material. Keeping those paths distinct makes the source of an answer clearer.
Explore the KiNDD architecture
Make AI fit the engineering workflow
At TECHNIA, I led technical delivery of Product Structure Navigator and ServicesIQ. I built an AI-powered Bitbucket PR reviewer and reusable coding-agent workflows, including browser verification of the resulting changes.
The surrounding delivery work mattered too: PSN’s CI and releases, and ServicesIQ’s CloudFormation, ECS Fargate, and RDS PostgreSQL infrastructure. A suggested change still has to survive review, checks, and release.
Build the systems around the model
For Stella at Ourself Health, I built the AI backend with agent orchestration, model routing, persistent sessions, and streaming responses. Its retrieval pipeline used pgvector with semantic and BM25 search to ground answers in source material.
I also built health-data ingestion, LLM-as-judge evaluations, and safety checks. Langfuse tracing and per-model cost tracking made the behavior and cost of requests visible. The backend ran on AWS infrastructure I built and maintained.
Stella and its supporting systems
Keep the source close to the answer
My Obsidian RAG system runs locally. It uses Python and ChromaDB to retrieve notes and Ollama to generate answers grounded in them. MCP exposes the same retrieval tools to my coding workflow.
I built Voicebox in SwiftUI to capture meetings and my microphone separately, produce transcripts, and prepare engineering handoffs. Those handoffs connect to the notes I search. I use both tools nearly every day.
How I built the local retrieval system
Current work
I’m a Forward Deployed Engineer on a consulting engagement. My background spans 14+ years of software engineering across enterprise systems, healthcare products, and cloud infrastructure.