A user's conversations and files live in their own Google Drive, and any connected model picks up the same context. Claude and Grok both connect through an MCP server I built.
The API, Postgres 17 and Redis run as containers on one Contabo host behind Cloudflare, with the frontend on a Cloudflare Worker. Moving the database next to the API cut the round trip from about 112 ms to about 5 ms.
Covered by 2,000+ backend tests and 480+ frontend tests, with live payments shipped through Paystack.
REZMO walkthrough — sign-in, workspace and the memory layer in use.
Three workflows in parallel: social posts to Facebook and LinkedIn five times a day, a YouTube video pipeline of up to two videos a day, and a job-application workflow that scans every 30 minutes, scores roles with Gemini and applies through Lever and Ashby.
A deterministic trigger layer (regex gate, then a Zod-schema extractor) keeps casual chat from ever firing a workflow. The LLM stack is Gemini with a Qwen2.5-7B model I host on Hugging Face as fallback.
It runs on free tiers (Render, Hugging Face) and cut the manual time I spent on these tasks by about 99.9%.
Recorded build session — wiring the job-application engine (Lever and Ashby submission path).
Honk is a sports prediction system that reports a 70–80% success rate. The engine underneath it is bookmaker-independent: it never reads odds, lines or market movement, scoring fixtures instead from raw kinetic and tactical match data.
Three structural pillars — kinetic velocity, defensive elasticity and structural integrity — feed a set of hard-coded framework patches, guarded by a market-decoupling veto and a virtual-fixture kill switch.
On finished fixtures it runs a Diagnostic Recurrence Analysis to autopsy why a prediction hit or missed, so the framework tunes from evidence rather than opinion.
Engine overview — the three pillars and the framework patches that guard them.
A non-custodial multi-chain wallet covering 15 chains, now decommissioned. It stays in this portfolio as an archived build: it is where I worked out key custody, encrypted transport and runtime hardening.
Private keys never left the device, and the backend only saw public addresses through an encrypted envelope (P-256 ECDH, HKDF-SHA256, AES-256-GCM) with the master key held on a separate server.
The operational patterns from that build — PM2 monitoring, iptables hardening, automated return-to-source — now protect the systems I run today.
CloudVoid at cloudvoid.online — the non-custodial onboarding flow, kept online as an archive.
Layout, motion, front-end build, hosting and DNS — usually on Cloudflare, Netlify or Hugging Face. These two are preview builds with no public domain yet, so the walkthroughs below are the demo.
A fashion house site built around a scroll-driven homepage: editorial lookbook panels, motion transitions and a fully responsive layout, backed by an admin console for products and content.
Designed and built end to end — visual system, page composition, motion, front-end build and deployment — with the admin panel shipped alongside the public site.
Walkthrough of the public site and the admin console. Preview build — no public domain yet.
A website for a therapy NGO: service cards for psychiatric care, counselling and community programmes, team and approach sections, impact reporting and a session-booking call to action.
Built as a responsive site with a matching mobile layout, designed for slow connections and older devices as much as for desktop.
Desktop and mobile walkthroughs. Preview build — no public domain yet.
An interface for the Eyeno second brain: a prompt workspace that diagnoses a draft across persona, objectives, delimiters and variables, then asks the user four alignment questions before it writes the final system prompt.
The knowledge behind it lives as Obsidian Markdown on GitHub rather than inside model weights, so the behavioural rules can be read, diffed and corrected.
Pipeline overview — the prompt checks and the vault nodes they query.
A containerised banking front end deployed on Oracle Cloud's Always Free tier. In production the heavy Node runtime is dropped for a static NGINX delivery layer, cutting live memory to under 5 MB. The image is multi-stage, so node_modules and build scripts never reach the running node, and the backend steps down from root to the node user. Kubernetes manifests live alongside the app.
The publishing agent behind the social workflow: a key-rotation engine pools API credentials and rotates on quota or invalid-key errors instead of crashing, while a dual-axis fallback cascades between models when a provider degrades. Memory is committed straight to GitHub over the REST API, because the host's disk is ephemeral.
Own-domain mailboxes served from the REZMO host over IMAP and SMTP, with SPF, DKIM and DMARC managed by me rather than delegated to a mail provider. Account mail, DNS and the product all sit under one operator.
Workers and Pages for front ends, DNS and TLS at the edge, WAF rules in front of origin APIs, and origin lockdown so the host only answers traffic that arrives through Cloudflare. Netlify and Hugging Face Spaces are used where a static deploy or an ML demo fits the job better.