Products
The consulting claim is stronger when the products are real. These are systems I designed, built, and in two cases still operate. The commercial source stays private; the architecture and the results are public.
A Chrome extension for job seekers that proves which jobs are real, finds the people who can refer you, and prepares your applications without inventing facts. Local-first data, on-device models where they fit, and a self-updating native helper.
Live with a free tier and three paid plans ($19, $39, $79/month). 16+ job boards, including LinkedIn, Indeed, Naukri, Workday, and Greenhouse. The server passes model requests through and keeps counts, not content.
A browser-native AI agent built into a Chromium browser. Tappi indexes the page, gives the model a reference to the element it needs, and sends only what the model needs — 3–10× fewer tokens than conventional browser tooling, with no navigator.webdriver footprint.
The product includes a Chrome sidebar that drives installed Codex, Claude Code, or Kiro harnesses; isolated agent tab groups; selected text and image snips; mobile emulation; editable design previews; and a bundled Office web add-in. The commercial source is private. The open-core SDK, architecture, and token accounting are public.
Also built
Products that are not for sale, but are part of the evidence: a multi-harness extension, an Android tutor, and the open-source browser utilities.
A Chrome extension and native host that drives Codex, Claude Code, and Kiro, with native messaging and an Office add-in. ~36.5K LOC and 69 test files. The sanitized architecture and extension shell are public; host paths, keys, and signing material are not.
Realtime voice tutoring, encrypted data, signed PDFs, and a GMS-free build. 632 Kotlin files, ~95K LOC, 522 tests at checkpoint. Client and independent product work. Student data and client details are sanitized.
Lightweight CDP browser control for AI agents. Connect to an existing Chrome session, control tabs and frames, and read pages with a small token footprint. MIT licensed and public.
The common thread
Every product above forced a production decision: install and update, local data, approval gates, cost per request, a failure that names itself. That is the same work I do inside a client's system.