Gaby
Zaynoun

The VAISS Compliance Auditor takes a plain-English description of an AI system and produces a board-ready compliance assessment against Australia's Voluntary AI Safety Standard and its 10 guardrails. Built as a multi-agent pipeline: a risk classifier determines the system's risk tier, a guardrail scorer assesses all 10 guardrails with calibrated green/amber/red ratings, and three generators run in parallel to produce an AI inventory entry, a board memo, and a prioritised 90-day remediation plan. Results stream to the interface in real time. The problem is concrete: every Australian enterprise running AI now faces 'are we VAISS-aligned?' from their board, and the current options are a $150k+ consulting engagement or a blank Word document. This tool produces a structured first-pass artefact in seconds — an assessment aid that gives a governance lead a starting position to react to, edit, and defend.

GabyGPT runs a stock Qwen3.5-4B with no hosted API anywhere in the request path. It runs on two devices from one shared core: the desktop over LM Studio, and Android with the model executing on the phone's own silicon — the phone half needs no desktop, no server and no network, so it works in airplane mode with the PC switched off. The model isn't the interesting part; the architecture around it is. Three things are kept deliberately separate and separately versioned: the system policy (authoritative, how to behave), persistent memory (facts and preferences), and the conversation. Qwen3.5 rejects a second system message, so policy and memory ride in one — policy first, memory explicitly framed as context that does not override it. That framing is the whole trick: without it, a remembered preference can quietly outrank a behavioural rule and the assistant drifts without ever telling you why. It is built to be measured rather than vibed. A frozen 41-case benchmark is the baseline, a separate memory benchmark tests recall, and a contamination gate keeps eval cases out of the training corpus. Corrections are captured in-session and reviewed by hand before they land, and forgetting a memory archives it rather than destroying it, so the memory state stays auditable. It is honest about what it is: the weights are stock, so the behaviour comes from policy, memory and context — not a fine-tune. Keeping that distinction sharp is what makes it possible to later measure prompting against an actually fine-tuned model instead of guessing which one did the work.

Pulling a table out of a PDF is trivial when the PDF is well-behaved and close to impossible when it isn't. This handles the cases that break naive extractors: unruled tables with no grid lines to anchor on, tables that continue across page breaks, merged cells, and scanned pages where OCR has to run before there's any text to parse at all. The output is a real .xlsx workbook rather than a CSV dump — one sheet per table, frozen bold headers, auto-filter, auto-fitted columns, and currency stored as actual numbers, so Excel can total a column while still displaying 'AUD 115,000.00'. The results grid behaves like a spreadsheet: drag-select a range, shift-click to extend, Ctrl+C pastes straight into Excel. Every cell carries a confidence score and low-confidence cells are flagged for review — because the failure mode that actually hurts with OCR isn't missing data, it's wrong data that looks right. It started as a way to take the tedium out of a repetitive data-entry task at work, and a few colleagues on the team use it now too.

CalcSolve is a production SaaS that solves calculus problems end-to-end and teaches the steps. The AI engine breaks each problem into named techniques (substitution, integration by parts, partial fractions, etc.) and walks the student through each step in natural language. Gamified skill trees give students a progression model, and Stripe handles tiered subscriptions. The challenge wasn't 'solve calculus' — LLMs can do that. It was wrapping the model in a UI that turns a one-shot answer into a teaching loop, with prompt structures that keep the explanations consistent and pedagogically sound.

An AI assistant built specifically for the dental vertical I shipped production software in at Stoneglass Industries. Multi-session chat memory with intelligent response orchestration, structured PDF export for clinical workflows, and Firebase/Firestore auth + data. The domain expertise from three years building CAD/CAM tools for dental manufacturing — including a TensorFlow 3D alignment model — informed every product decision. This is what 'AI for a vertical' looks like when the developer actually understands the vertical.

Tileverse is a production marketplace I built solo end-to-end: Next.js frontend, Supabase for data + auth, Stripe for payments, and an anti-abuse system to prevent automated grid claims. The hard problems were the economic and trust ones — pricing curves for scarce digital assets, abuse-resistance without killing user experience, and a payments flow that handles both upfront purchases and ongoing customisation.

TradieSpark is a productised service business: brand, website, pricing, onboarding, and outbound sales — built and operated solo. Tradies need web presence but can't afford or evaluate a custom agency engagement. The product is a fixed-scope, fixed-price website with predictable delivery, sold and delivered with AI-assisted development. It's as much about scoping and pricing as code.