Skip to content

Selected work

The problem, the decisions, and what shipped.

Client names and identifying detail are removed at our clients’ request. Everything else — the constraints, the architecture, the tradeoffs we made and why — is exactly as it happened.

01

2026

Legacy Modernization · Nonprofit / Member Operations

Community Operations Platform

Two aging Web Forms applications, rebuilt as one portal that a volunteer committee can actually run.

Applications consolidated
2 → 1

Applications consolidated

Functional domains specified
12

Functional domains specified

Production infrastructure cost
~$13/mo

Production infrastructure cost

Database passwords in circulation
0

Database passwords in circulation

.NET 10ASP.NET CoreEF CoreAzure SQLAzure App ServiceAzure FunctionsEntra IDManaged Identity
Read the case study
02

2026

AI Engineering · Creator Economy / Applied AI

Grounded Voice Agent

A real-time voice assistant that answers in a founder’s own words — because it is required to look them up first.

Real-time voice, interruptible
Full-duplex

Real-time voice, interruptible

Local embedding index, no vector vendor
384-dim

Local embedding index, no vector vendor

Tools available to the agent
4

Tools available to the agent

Answers preceded by retrieval
100%

Answers preceded by retrieval

PythonLiveKit AgentsGemini Realtimesentence-transformersscikit-learnJSONL corpusTool calling
Read the case study
03

2026

Multimodal AI · R&D / Wearable Computing

Smart Glasses Field Assistant

Hands-busy work, answered out loud. Live camera and voice from wearable hardware into a multimodal model.

Surfaces: glasses, phone, desk
3

Surfaces: glasses, phone, desk

Wake word — never always-streaming
On-device

Wake word — never always-streaming

Spoken answer budget
1–2 sent.

Spoken answer budget

Face recognition, no cloud enrollment
Local

Face recognition, no cloud enrollment

SwiftSwiftUIFastAPIGemini Live APIopenWakeWordONNX RuntimeOpenCVMediaMTXWebSockets
Read the case study
04

2026

Product Engineering · Education / Consumer Product

Offline-First Study Platform

An 8,000-entry classical dictionary turned into a study product — web and native, shipped, and fast offline.

Dictionary entries indexed
8,384

Dictionary entries indexed

Store releases shipped
20

Store releases shipped

Downloaded once, then never again
7.6 MB

Downloaded once, then never again

Search latency after migration
Seconds → instant

Search latency after migration

React 19TypeScriptViteExpoReact NativeSupabaseIndexedDBts-fsrsEAS Build
Read the case study

Next step

Tell us what is stuck. We will tell you what it takes.

The first conversation is a scoping call, not a pitch. You leave it with an honest read on feasibility, sequence and cost — whether or not we end up working together.