Book a free call
Selected Work

Real AI automation, built and running.

Operator-built, finance-grade, human-approved AI automation. Below: systems I built for a North American back-office services firm, then products and systems I have built on my own, from a voice-first assistant to a shipped learning marketplace.

How to read this

One pattern runs through all of it: AI reads, code decides, a person approves.

Each build below reads messy real-world inputs with AI, hands the decisions that matter to deterministic code, and stops for a human to approve anything that changes a record. That is what "finance-grade" means here: nothing posts on its own.

Everything below is grouped in two. First, AI automation I built for a North American back-office services firm, anonymized where client confidentiality applies. Then systems and products I have built on my own, where there is no client to protect. Either way, no client is named, sensitive domains stay general, and every diagram and mockup is a schematic with placeholder numbers, never a real screen or real client data.

Finance-grade Human-approved Runs inside your own systems Business Central, Azure, React, FastAPI, Snowflake
Part one · Client work, anonymized

AI automation for operations

Systems I built for a North American back-office services firm, anonymized where client confidentiality applies. Every number in every diagram is a placeholder.

01 / 08
Billing automation · back-office services firm

Contract-first billing engine with a self-improving accuracy loop

For a North American back-office services firm, I built the engine that assembles its entire monthly client invoice book automatically, pricing every line from the signed contract rather than a drifting spreadsheet and flagging anything uncertain for a human.

I then added a self-improving accuracy loop that grades its own output against the approved books and keeps only changes that raise the match, which recovered tens of thousands of dollars a year in quietly under-billed revenue.

~$45K / yrof quietly under-billed revenue recovered, at about 95% of drafts matching the reviewed book

One firm's outcome, not a typical or guaranteed result.

Usage data Snowflake Contract -first pricing QA grade Draft book Business Central Self-improving accuracy loop Tunes itself offline. Never touches production. Candidate change Score vs approved close Keep only if better Rejected changes are discarded. Only improvements ship. ~95% match · tens of thousands recovered / yr
Schematic flow. No real invoice data shown.
02 / 08
Human-in-the-loop AI · back-office services firm

Voice-guided AI close agent and console

For a North American back-office services firm, I built a voice-guided AI agent that runs the firm's monthly client billing close alongside a person. It checks the data is ready, drafts every record, and proposes each system write as a card the operator approves by voice or click, with nothing posted on its own.

It runs inside the firm's own systems and turned a long manual checklist into a three-step guided flow used by non-technical staff.

~95%match to the manual close, now behind a plain-language, approve-every-step interface

One firm's outcome, not a typical or guaranteed result.

Close Console Ava · ready 1 · Sources 2 · Review & Edit 3 · Finalize Ava narrates each step Approve this write? Draft only · nothing posts on its own Draft invoice · Client 0000 Amount $0,000.00 Go ahead Edit
Schematic mockup. Approve by voice ("go ahead") or click. Numbers shown are placeholders.
03 / 08
Data migration · back-office services firm

Compliance-document migration pipeline, at scale

For a North American back-office services firm, I built a pipeline that moves a client's compliance and personnel documents from a prior provider's system into the new one during onboarding, work that used to be done by hand file by file.

It reads the old system's API, re-labels and sorts every document to the new system's rules, and packages it for a one-shot upload. One migration handled several thousand documents at around 99% success, in hours rather than days.

~7,400documents moved in a single migration, at about 99% success, hours not days

One firm's outcome, not a typical or guaranteed result.

Prior system API Read every document Re-label & sort Package New system Manual vs automated Manual days of work, file by file Automated 2 to 4 hours, end to end Around 7,400+ documents across ~648 employees in one run, at ~99% success.
Schematic. Volumes shown are from one migration; no client data displayed.
04 / 08
Support AI · back-office services firm

Support AI agent and voicemail-to-ticket pipeline

For a North American back-office services firm, I built an AI layer over its client support desk: it drafts the first reply from an approved knowledge base for a human to send, and it refuses to answer anything sensitive on its own.

I also built a live pipeline that catches every missed-call voicemail, transcribes it in seconds with cloud speech AI, and drops the text straight onto a tracked support ticket so nothing gets lost, for about a dollar a month.

~45%of inbound calls answered at baseline; every missed voicemail now becomes a tracked, transcribed ticket

One firm's outcome, not a typical or guaranteed result.

Missed call voicemail left Webhook fires Transcribe Azure Function + AI Speech Support ticket ~2.7s / voicemail Ticket #0000 · New Voicemail Transcript note (auto-added): "Hi, this is a sample caller about ..." Suggested reply drafted - a person approves before it sends.
Schematic. Ticket text and numbers are placeholders; nothing sensitive is auto-answered.
Part two · My own builds

Systems and products I have built

Independent builds of my own, with no client involved, from a voice-first assistant and a walled-off multi-agent system to a shipped learning marketplace and an AI video pipeline.

05 / 08
Voice-first assistant · operator's own build

Jarvis, a bilingual voice-first AI assistant with a live HUD

I built Jarvis, a bilingual English and Spanish voice-first AI assistant with a live heads-up display. I talk to it hands-free and it gives me a morning brief, project status, and upcoming deadlines, captures notes and rules by voice, and shows live widgets, with an AI model as the brain and my own intent router pre-loading the right context for each request.

Anything that writes is read back and confirmed first, so nothing is saved without my okay. Built with Python, faster-whisper for speech-to-text, neural text-to-speech, a browser HUD over WebSocket, and a custom intent router with confirm-before-write.

Hands-free · bilingual · confirm-before-writeEverything it saves is read back and confirmed first, so nothing is stored without my okay.

One operator's own build, not a typical or guaranteed result.

JARVIS bilingual voice HUD · EN / ES LISTENING JARVIS // LISTENING intent router · context pre-loaded VOICE TRANSCRIPT >> what's my morning brief? << brief ready · 2 deadlines this week >> capture a note for later Capture this note? Read back first · saved only on your okay Confirm Not yet LIVE WIDGETS WEATHER 00° cloudy · sample MARKETS ▲ 0.0% index · sample NEWS
Schematic HUD. Widget values are placeholders; anything that writes is read back and confirmed first.
06 / 08
AI architecture & safety · operator's own system

Multi-agent AI assistant with a walled-off front desk

I designed and run a multi-agent AI assistant on a small always-on server, reachable from a phone, that can send email, manage a calendar, and handle files, with a hard rule that only its owner can trigger any action that reaches out or changes something.

It includes a separate front-desk agent for outside contacts whose safety is built into its structure rather than a promise: it has no access to private tools at all, and it can confirm meeting times through a calendar view that shows open slots but never the details.

0private tools reachable by the outside-facing agent; containment held against an "ignore your instructions" attack

One operator's own build, not a typical or guaranteed result.

Private agent Machina · owner only Machina Gmail send Calendar Files Only the owner can trigger outward or destructive actions. HARD WALL Front-desk agent Iris · outside contacts Iris Messaging only No shell, no code No files, no secrets Free / busy only · titles hidden
Schematic. The wall is structural: the front-desk agent has no private tools to reach, verified in the runtime trace.
07 / 08
Shipped product · operator's own build

Pocketclass, an online learning marketplace

I designed and built Pocketclass, a full online learning marketplace, as an independent product. Instructors publish courses and live sessions, students browse and buy, and an admin oversees the whole thing, with real sign-in, Stripe payments, and a commission-based pricing engine.

It runs on a modern web stack, Next.js, React, Supabase, and Stripe, and ships a demo mode that works entirely on sample data, so it can be shown safely without touching a real database.

Shipped product · real paymentsA complete, multi-role marketplace with real sign-in and Stripe checkout, shown safely here in demo mode on sample data.

One operator's own build, not a typical or guaranteed result.

Student · Catalog Sample Course A $00 Sample Course B $00 Buy · checkout sample data Instructor · New course 1 2 3 Title Sample title Price $00 Lessons 3 Revenue (sample) Publish course Admin · Overview Users 000 Courses 00 Payouts $0,000 Sales (sample) demo mode Stripe checkout 3 roles: student · instructor · admin
Schematic mockup in demo mode. Sample courses and prices are placeholders, never a real user's data.
08 / 08
Media automation · operator's own build

AI-driven video editing pipeline

I built an AI-assisted video editing pipeline that turns raw footage into a finished cut without a manual timeline. It extracts the audio and frames, transcribes them, and asks a model to choose the strongest clips and write an edit list, then renders the final video programmatically in both wide and vertical formats.

It turns a manual editing session into a repeatable, code-driven process a person still reviews. Built with Node.js, Remotion, faster-whisper, and ffmpeg.

AI plus code · finished cutA model chooses the clips and writes the edit list; code renders the finished cut in both wide and vertical formats.

One operator's own build, not a typical or guaranteed result.

Raw footage Extract audio + frames ffmpeg Transcribe faster-whisper Model picks the clips + writes the edit list (JSON) Render Remotion 16:9 highlight reel 9:16 story reel
Schematic pipeline. Thumbnails are neutral placeholders, never real footage; a person reviews the final cut.
The method

Why these builds pass a security review.

Drafts, not auto-posts

The system prepares the work and proposes each change. A person reviews and approves anything that matters. Nothing runs on its own.

Deterministic where it counts

AI reads and drafts; plain code does the math and the rules, so the same inputs always give the same answer and results are auditable.

Inside your own systems

Built to run in the tools you already have, like Business Central, Azure, and Microsoft 365, with least-privilege access and a security one-pager for your IT team.

Results described on this page are one operator's or one firm's outcome and are not a typical or guaranteed result. Every visual is a schematic mockup with placeholder numbers, not a real screen or client data.

Want work like this in your business?

Start with the AI Automation Audit: fixed fee, fixed timeline, no obligation.

Book a free 30-minute call