Hi, I'm Yousef.
I build software for real teams.

Personal project · Oct 2026 · live

Nightshift

A night shift of AI bots for small businesses. While the owner is off, the bots read the inbox, flag what matters, draft replies, chase invoices, follow up on leads and do research, then leave a morning brief on the owner's desk. Nothing that sends an email or moves money happens until the owner approves it. It's a web app and a command-line tool, running on DeepSeek, Supabase and Vercel, with live Stripe billing at nightshiftai.ca.

## The office

The main screen is a pixel-art office. Every bot has a desk and a status you can read at a glance: typing while it works, a sign saying "2 drafts for you" when something is waiting, z's drifting up while it's idle until its next shift, a wobbling "I'm stuck." when a connection it needs is missing. The office lays itself out in balanced rows and scales to fit any window, so it never scrolls. Click a bot to see its last runs and what it's waiting on; click your desk for the morning brief and the numbered queue of cards.

The bot room is where the team works in the open: one bot hands a job to another ("@Rosa, Anika says she was charged twice"), @mentioning a bot gives it a job, and holds show who the team won't contact and why. Each bot also has its own chat, where its drafts appear inline with Approve and send, Edit, Ask for changes and Discard. On a phone it all becomes a list, an approvals screen and a tab bar.

The bots are cute on purpose. They hop with a little heart when you approve their work, the team waves when your desk is clear, the room's first day has three heads bobbing, and buying a plan ends with the whole team hopping under pixel confetti. Every animation is under a second, never blocks anything and turns off with the system's reduce-motion setting.

## Made for each kind of business

There are 424 bots to hire, most of them specialists for one of 101 kinds of business: recall and chair-filling for dentists, no-heat dispatch for HVAC, deposit chasing for caterers, intake for immigration lawyers. Each industry also gets its own cast for the five core jobs (inbox, research, billing, follow-ups, the brief), with their own names, job titles and words. A dental practice's inbox bot is "Mirela · Front desk", who answers patients, books appointments and treats tooth pain as urgent; an HVAC company's is "Dex · Dispatch", who treats no heat or a gas smell as urgent. The bots hand work to each other by those names, the brief has the industry's own sections ("Today's chairs", "Today's service calls"), the chat suggestions use its words, and the app's copy follows. Change the industry and the team takes the new names, except any the owner renamed.

## Nothing goes out without your OK

Every tool has a risk class. Quiet tools (reading, flagging, labelling, drafting) just run. Message tools (sending an email) turn into a card the owner approves, unless they've told that bot it may send on its own. Money tools (invoices, credit notes) always wait, whatever the setting. Approving starts an 8-second undo, then the card goes out.

• Every approval step is a guarded move: it only happens if the card is still where it was, so undo can't race the send and nothing goes out twice.
• Money is checked three times (when the bot drafts it, when the owner edits the card, before it runs). A credit note can't exceed its invoice, and approving money needs the authenticator step on every path, including "Approve all" and the CLI.
• Before anything outward, a guard checks the do-not-contact list, any holds (a client disputing a bill pauses their reminders) and a weekly cap on follow-ups. It also refuses recipients that could smuggle in a hidden Bcc.

## How the bots work

Every bot runs on DeepSeek with native tool calling: Flash for the everyday work and Pro for the big write-ups. Each one gets a system prompt built from its job, the business, the owner's settings, a shared team memory and who its teammates are, plus only the tools for what's connected (35 in all). A tick runs every 5 minutes from Supabase's pg_cron. It sends cards whose undo has passed, runs queued hand-offs, retries failed runs and starts every bot whose shift is due. It's built to survive a bad run or a slow model, and it clears anything a crashed tick left behind. Each workspace has a daily spending cap on model and search costs, and Dawn's brief arrives at the owner's chosen time by email, push and in the app.

## Connections

Gmail and Google Calendar, Outlook and Microsoft 365, QuickBooks, Xero, FreshBooks and Stripe for the books, Shopify for orders, HubSpot, Google Sheets or a CSV for leads. The OAuth apps are all set up and verified. Tokens and keys are encrypted with AES-GCM before they're stored, and refreshes are safe when two happen at once, even with providers that hand out a new refresh token every time.

## The business around it

Billing is live on Stripe: Solo, Studio and Team, monthly or yearly, with extra seats. Switching plans is prorated, cancelling happens at the end of the period, and the owner can take a cancellation back, all from inside the app. A purchase shows up even if a webhook is late, because the billing page reads the subscription straight from Stripe. Nightshift's own email goes through Resend from nightshiftai.ca: sign-in codes, the morning brief and alerts to me about payments, failed payments, plan changes and new sign-ups, all in the same design. An admin view shows every workspace, user, failed run, spend and revenue, read-only and behind the authenticator step.

The CLI signs in with a device code (the approval page shows which device and network is asking), then lists the queue, approves cards by number, follows the bot room live and gives bots jobs from the terminal.

It's covered by 479 tests. Two agent audits went through the backend and the UI, and their findings were fixed with tests to match. The screenshots here come from a showcase mode built for them: on the local dev server, the real screens fill with a made-up practice generated by the app's own server code.

• 424bots to hire
• 101industry casts
• 35tools
• 479tests passing
• 56API routes
• 8 sundo on every send

• Next.js 16
• React 19
• TypeScript
• Supabase
• Postgres + RLS
• pg_cron
• DeepSeek API
• Tool calling
• Stripe Billing
• Resend
• Web Push
• OAuth 2.0
• Gmail API
• Microsoft Graph
• QuickBooks
• Xero
• FreshBooks
• Shopify
• HubSpot
• Server-sent events
• Vitest
• Vercel
• CLI

The practice, the patients and the people in these screenshots are made up.

Personal project · Oct 2026

yAI

A ChatGPT-style app and coding agent built on models I wire together myself: four hosted models that answer right away, a coding model on a cloud GPU that only exists while someone is using it, and a small model on my laptop's GPU.

## Models

The site is on Vercel. The default, yAI Lite (GLM-4.7-Flash on Z.ai), is fast and free, with DeepSeek V4.1 Flash behind it if Z.ai is busy. yAI Ultra Flash (DeepSeek V4.1 Flash) costs a fraction of a cent per reply, and yAI Ultra (DeepSeek V4 Pro) handles the hardest questions. Both run on DeepSeek's own API, where the repeated start of a prompt is cached and billed at a thirtieth to a fiftieth of the normal price, and everything is half price outside DeepSeek's peak hours; Kimi K3 on Modal takes over for yAI Ultra if DeepSeek is unavailable. yAI Q3.8 Max (Qwen3.8 Max, 2.4 trillion parameters) is a strong all-rounder on Modal. Anything with an image goes to free vision models instead: Gemini 3.5 Flash-Lite first, then Qwen 3.8 27B on Groq, then Gemini 3.8 Flash, each one handing over the moment its free quota runs out. Every reply's cost, at the price of whichever host answered, is added to a daily allowance per person in Upstash, so the bill can't run away, and admins can raise or lower each person's allowance.

A small model, yAI Q3.5, runs on my laptop's 6 GB GPU at up to 200 tokens per second: each question travels through a Cloudflare tunnel to a token-protected bridge in front of LM Studio, and it appears in the model menu whenever my PC is on.

It also makes images and videos on RunPod's serverless ComfyUI GPUs. Qwen-Image 2.1 creates pictures from a description or edits a photo you upload, and MiniMax H3 turns a prompt into a short video with sound. Everything goes through a safety filter and daily limits. Accounts, saved chats, images and videos are stored in Upstash Redis, so everything works even when my PC is off.

## The web app

The site has its own design with matching light and dark modes: warm neutrals, Geist type and one purple accent, and a sidebar that folds into a rail. On a phone it becomes an app that opens on the logo, with Chats, Projects and You tabs, bottom sheets for the model and the tools, and swipe-back navigation. A Search button looks the question up on the web, reads the top pages, and answers with numbered citations that link to each source. Projects keep files, PDFs and instructions together, so every chat started in one already has them. HTML, SVG and JavaScript in an answer open in a live preview that runs in a sandboxed frame. When a picture explains it better, answers include diagrams (flows, sequences, timelines) and charts drawn from the numbers, each with a table view and a download button. A usage page shows each person their messages, images, videos, searches and spend on the hosted models against the daily limits.

Deep research plans several searches, reads up to about fifteen sources over two rounds and has yAI Ultra write a cited report. Data analysis loads a CSV, Excel or JSON file into Python (Pyodide) right in the browser, so the file never leaves the device; the model writes pandas code, it runs there, and the charts appear under the answer. Memory keeps lasting facts you mention and lets you edit them. Voice mode takes dictation, reads any answer aloud, and holds a hands-free conversation. The canvas opens long writing and code in an editor beside the chat, with versions, "select a part and ask yAI to change it", and live HTML, JavaScript and Python runs. Ctrl K searches every saved chat and jumps to the exact message.

Answers can also call connected services for live, exact data: the weather, maps and directions, stock and crypto prices, sports scores, Wolfram Alpha, arXiv and Semantic Scholar papers, YouTube transcripts, your GitHub repositories and issues, UWaterloo LEARN due dates and course info, and webhooks to Discord, Slack or n8n. Live results show as cards, and watchers send a notification when a price drops or a page changes.

For school there's a study mode that tutors with hints and questions instead of answers, and a Study page that turns notes, a PDF or a chat into flashcards with spaced repetition and practice quizzes, with short answers graded by the model. A calendar and email tool reads a connected Google account and suggests Gmail drafts and calendar events as cards that are only saved when you press the button. Made images get an Edit button, one-tap edit ideas and a before/after slider, and the whole site installs as an app on a phone or laptop. Scheduled tasks run a prompt on their own every day, on weekdays, weekly or once (a morning news briefing, say), save the answer as a chat and send a notification to the phone.

Around it I built the parts a real service needs: an OpenAI-compatible public API with personal keys, counted against each person's limits; an admin People page with everyone's usage, per-person limits, account pausing and automatic flags for unusual use (bursts of requests, the same message over and over); and a What's new page where every release is numbered by its build.

## yAI Code

yAI Code brings the same models to the terminal: a coding agent like Claude Code that signs in with your yAI account and works inside your project folder. It reads and searches the files, edits them in place, and runs shell commands to build, test and lint the code. By default it asks before every file change and command, so nothing happens until you approve it. On DeepSeek it calls its tools natively, and it never reads the same lines of a file twice when they're still in front of it.

It also gets a model of its own: yAI 3 Coder (Qwen3 Coder 30B), a coding model with a 64K-token context and the one model still on a dedicated GPU. It runs on a 48 GB L40S on Modal with vLLM, keeps its weights on a Modal volume so a cold start loads them instead of downloading them again, and shuts down to zero when idle. It's a mixture-of-experts model that uses only 3 billion of its 30 billion parameters for each token, so it's quick for its size.

I benchmarked it inside yAI Code on six small Node.js tasks: fixing bugs, writing an LRU cache and an async retry helper from empty stubs, rewriting a CSV parser, and adding features across two files. Every task starts with failing tests, and the score is how many pass afterwards. yAI 3 Coder passed all 44 without touching a test file, and the simpler tasks took about 20 seconds each.

• 44/44tests passed
• 6/6tasks solved
• 0test files changed

It has three permission modes: ask prompts before any edit or command, auto-edit applies file changes automatically but still asks before running commands, and plan is read-only, so it explores the codebase and proposes a step-by-step plan without touching anything. You can @-mention specific files to point it at the right place, and each project folder keeps its own saved conversations, so picking up a project later resumes right where you left off.

It also handles the rest of the coding loop. /fix runs the tests and keeps fixing the code until they pass, and /fix watch reruns them on every save. /review has a second model review the uncommitted changes, and /commit writes the message and commits. It can see, too: /look screenshots the page on desktop and phone for a design review by a vision model, an @-mentioned mockup image becomes something the coder can build from, and /showcase records a demo video of the page. In /duo mode two models work together, one planning and reviewing while the other writes the code; /tutor gives hints instead of finished solutions, /share publishes a session at an unlisted link, and /undo reverts its last edit.

It runs as a full-window app so you always know what it's doing, what it needs and what you can press: a header on top, the conversation in the middle with the plan in a panel beside it, and the input box and keys at the bottom. A status pill says Ready, Working, Finished or Needs attention, approvals say what a command does and how risky it is and take a single key, and a finished task ends with a summary of tests, files and cost. It searches the web when it needs documentation, runs dev servers as background tasks, connects to MCP servers, and comes as a VS Code extension that opens every proposed edit as a diff. yai remote sends a task from my phone to my PC, and /clone rebuilds a page from a screenshot or a URL, then compares its result with the original and fixes the differences.

## /team and /crew

/team runs a whole team of agents. A lead model splits the task, a researcher looks up what a great version needs, an architect writes a contract of every shared name and a skeleton of every file, and up to eight builders work at the same time, each owning a feature. After them come a merger, yAI's own browser checks (it opens the page, gets past its menus, measures the frame rate, checks phone sizes and plays games for 30 seconds), a single QA agent, a fixer working from one merged to-do list, and a spec checker that scores the result against the request. A judge plays it, and the next round improves on what it saw, with a pause between rounds to try it yourself and say what to change. A watcher reads every agent's steps every 20 seconds, keeps their to-do lists accurate and redirects one that's going round in circles. A live board shows each agent's step, time and cost, and every run ends with a report page: a replay of the board, the score, screenshots, and a table of where the time and tokens went.

/crew came out of what /team taught me: on a game, one agent that holds the whole picture beat a team of builders working apart. So in /crew a single main builder builds the whole project itself, checking the page as it goes, and commissions helpers for the separate jobs (level data, sprites, sounds, tests, research) on yAI Lite or Ultra Flash; they work at the same time and report back to it. The same checks, fixes and rounds run after it. In the video, /team's fourteen agents built a Windows 95 style desktop with six apps from one prompt.

• JavaScript
• Node.js
• CLI agent
• Multi-agent
• Vercel
• Modal
• DeepSeek API
• Z.ai GLM
• Gemini API
• Groq
• Qwen3.8
• Kimi K3
• RunPod
• vLLM
• ComfyUI
• Upstash Redis
• MCP
• VS Code extension
• Headless Chrome
• Pyodide
• OpenAI-compatible API
• PWA
• Cloudflare Tunnel
• LM Studio

yAI is invite-only. The hosted models answer right away; yAI 3 Coder, images and videos can take a minute or two to start after a quiet stretch while their GPU wakes up.