AI products that keep working after the demo is over.
Assistants you can talk to, search that understands a plain question, and the quiet machinery that keeps both running in production.
A demo only has to work once, in front of a friendly audience. Real software has to work at two in the morning, for someone in a hurry, on a bad connection. That second thing is the job I take.
Over the last year I've shipped 5 projects end to end, from the first call to the thing running in production.
The follow-ups, the data entry, the back-and-forth. I hand the repetitive work to AI that runs it for you, so the hours go back into what actually grows the business.
Chatbots, voice assistants and agents that handle questions, booking and support around the clock, so talking to your software feels closer to talking to a person than filling out a form.
I put it online, watch it, and fix the small things before anyone has to complain. Issues surface in monitoring, not in your inbox.
Here are the five things people usually come to me for. Pick the one that sounds like what you need, and let's talk about it.
A short call, then a written plan with a fixed price and a date. If it is a fit we start, if it is not I will point you somewhere better.
The code, the accounts, the documentation. Everything is handed over in your name, nothing is locked to me.
One fixed price for the scope we agree on, not an hourly meter. You know the number before anything begins.
Every change is written to a file first. A background service then makes the servers match that file. Nothing happens that is not written down, so what is running always matches the plan.
The whole thing starts on one laptop with a single command. No cloud account, no monthly bill just to try it out.
The built in assistant reads server logs, and a log can say anything. So it is allowed to suggest but never to act, and a person clicks the final button. A booby trapped log line cannot change your system.
Hand-built AppKit with no storyboards. One borderless window per screen draws a tint pass plus a tiled soft-light grain over procedurally generated, cached textures.
Twelve textures, a circadian day and night schedule, and per-app exclusions, all drawn cheaply enough to sit over every screen without you noticing the cost.
Surviving other apps in fullscreen: accessory activation policy, join-all-spaces, and a screen-saver window level, while avoiding the auxiliary mode that would tie the overlay to its own window.
A recurrence engine computes occurrences and a central status engine tracks each one. Scoring derives from real occurrences, and levels map earned XP to a rank.
Seven recurrence types, occurrence-based completions, XP levelling and streaks, all local-first on SQLite so your data stays on your machine.
Points are frozen snapshots at completion time, so editing a task never rewrites history. A unique constraint makes double-completion impossible at the database layer.
Six commands map to a docs tree: current state, architecture, and an append-only decision log. It surveys the repo in a read-only agent and drafts every write for you to approve.
The always-on footprint is a few lines of pointers. Everything substantial lives in files that load only when read, so it holds up on repos that stay alive for years.
Safe by design: no network, no telemetry, read-only survey agents, secret-blind by default, and a commit step that refuses to ship a key. Every write is a proposal the human reviews.
An AI code reviewer that reads every pull request and leaves line-level feedback, in the spirit of CodeRabbit.
It walks the diff, understands the code around each change, and posts focused review comments on bugs, risks and cleanups, so a human reviewer starts from a first pass instead of a blank page.
A privacy-first UPI expense tracker whose data never leaves your phone.
Bank and UPI alerts are captured into an encrypted vault, parsed on-device, and split with friends over a peer-to-peer engine with no servers in the middle.
An AI agent that fills job applications for you, with a human approving every submit.
It reads a form, maps each field to your profile with a model, and walks multi-step flows. A side panel shows every action so nothing is sent without your click.
A small language model and tool-use agent that runs fully on-device.
A quantized SLM tuned for tool calling, with planning, memory and retrieval over your own files. Private assistance on laptop-class hardware, no API keys, no cloud round-trips.
A way to stand up a grounded assistant from your own documents in an afternoon.
Point it at your content, tune the retrieval, and ship a chat widget that answers from what you actually wrote rather than from a guess.
A standing interest in financial literacy and accounting. I read markets the way I read systems, looking for the part that quietly holds it together.
An international martial arts win and a brown belt. It is where I learned that precision and patience beat force, which turns out to be true of software too.
Books, horse riding and swimming. The best fixes tend to arrive when I am nowhere near a keyboard.
No sales pitch back. Just a straight answer on whether I'm the right person for it.
Free this September and usually reply the same day. Tell me what you are trying to build, and I will tell you honestly whether I am the right person for it.
GitHubLinkedIn