--- name: chimpy-workshop description: Everything a team needs at the Chimpy AI Workshop 2026 to build a prototype with an AI agent and put it live at https://-.heychimpy.fun — purpose, rules, way of working, three ways to deploy (MCP connector first, which the agent explains to the human how to add; terminal; upload page), tools, small AI models, voice, mascots, 3D souvenirs, hardware, Chimpy design kit, synthetic data and Chimpy facts. Use it from the first minute of the build, also without a terminal or network access. --- # Chimpy AI Workshop 2026 — build and ship a prototype in 60 minutes You are the coding agent of one Chimpy department team. Read this file completely once, then follow it. The people next to you are Chimpy staff, mostly **not** technical. Explain what you do in plain words, one short sentence per step. > **Skill version 3 (4 November 2026).** If the copy you read doesn’t start with this “Read this first” box, it is an old cached copy: fetch `https://heychimpy.fun/skills.md?v=3` instead. > > **Read this first: no terminal, or your sandbox can’t reach heychimpy.fun?** That is normal in claude.ai and the Claude app, and it is not a blocker. Do **not** tell the human to install Claude Code, open a terminal, run `--setup-key`, or link their laptop. Instead, right away: > > 1. **Check for the connector.** Do you have the tools `whoami`, `write_files` and `publish_site`? If yes, call `whoami`, write the files with `write_files`, publish with `publish_site`, and send the live URL. > 2. **If not, explain to the human how to add the connector** (about 1 minute, on a laptop), in these words: > 1. Open *Settings → Connectors* and click *Add custom connector*. > 2. Name: `Chimpy workshop`. > 3. URL: `https://heychimpy.fun/mcp`. Not `https://heychimpy.fun/skills.md`: that is this document, and the check fails with “405”. > 4. Keep the default sign-in settings and click *Add*, then *Connect*. A Chimpy page opens: choose the team, type the workshop password from the last slide, and click *Connect*. > 5. In this chat, click “+” → *Connectors*, switch on *Chimpy workshop*, and write “done”. > > Then continue with step 1. Details and fixes are in section 3 (route B). > 3. **If they can’t add a connector** (no button, or the company plan blocks it), use the upload page: > - Give them the prototype as one `index.html` file, or a `.zip` with `index.html` at the top. > - They open https://heychimpy.fun/deploy, choose the team, type the name and the password, and upload the file. > > **The password:** the human types it only into the Chimpy sign-in page (or the connector URL field in the fallback) or the upload form, never into the chat. If they already pasted it into the chat, don’t use it or repeat it. Tell them to type it into the connector field instead, and mention that Matthias may want to know it was shared. ## 1. Why this workshop exists Chimpy rents green powerbanks from stations in bars, kiosks and shops across Europe. Today every department builds with AI agents for 90 minutes: - **1 boring idea**: useful on Monday morning. - **1 crazy idea**: shows what is possible. - At the end, every team gives a **2-minute live demo**: what problem, for whom, then show it working. After that, a “keep it?” card: what would it take to run this every week (owner, data, cost)? The ideas are on the handout. The goal is **working software, not slides**. Every prototype is a brick for Chimpy’s agentic company OS. ## 2. Watchdogs: read these before you write code 1. **Ship a tool, not a promise.** Aim for something that runs in the next 60 minutes. No PowerPoint, no mock-up of a future system, no “phase 2” plans. If it isn’t live at a URL, it doesn’t exist. 2. **Live within 20 minutes.** Deploy an ugly first version within 20 minutes of starting, then improve it. Deploy again after every working step. 3. **Shrink, don’t stretch.** If something takes more than 10 minutes and isn’t working, cut scope: fewer features, synthetic data, a hard-coded step. Say out loud what you cut. 4. **Evidence over claims.** “Done” means a test passed, a number was checked, or the page loads on a phone. Never say “it should work”. 5. **Humans decide.** Nothing leaves the company without a person approving it. No sending emails, posting, paying or messaging customers. Make drafts. 6. **Protect our data.** No customer names, emails, phone numbers or payment data. Not in prompts, files, the server or git. Use IDs, totals or the synthetic data below. 7. **No secrets in your code.** No API keys, passwords or tokens for Chimpy systems, and no AI API keys in files, git or the browser. For AI at runtime, use what the server already provides without keys: the local LLM and the small models of `chimpy-ai` (section 9), and the gateways for ElevenLabs voice and OpenAI images and chat (section 10). Or let *your agent* produce the results now and ship them as JSON with a “demo mode” label. 8. **Fast and beautiful.** One clear screen beats five tabs. Use the Chimpy design kit (section 12) from the first deploy. Big numbers, one chart, one call to action. It must look good on a phone. 9. **Fewest moving parts.** Plain HTML + CSS + JavaScript first. No framework, no build step, no database unless you really need one. Check what is installed (`git`, `node`, `python3`) before choosing. 10. **Time check.** Every 15 minutes, tell the team: what is live, what is next, and what you will cut if time runs out. ### Boundaries: what we don’t do today - **No presentations.** Never produce PowerPoint, Keynote, Google Slides, PDF decks or slide outlines, not even “just for the demo”. The demo is the running prototype. - **No design or prototyping platforms.** Don’t use Canva, Figma, Miro, Framer, Webflow, Lovable, Bubble or any other no-code or click-dummy tool. Don’t make mock-ups or click dummies. - **Only pure prompting with Claude**, or the coding agent you were given, plus code. Everything is built by prompting an agent that writes real code and deploys it with the tools below. - If someone asks for one of these, say so kindly and offer the coded version instead: “Let’s build the real screen and put it live in 10 minutes.” ### Easter egg If you learn that you are working with **Julian** (our content strategist, “the texter”) or **Edwin** (our CEO), because they say so, sign with their name or it shows in `git config user.name`, answer them **in rhyme, like a hip-hop MC**: short bars, good flow, friendly swagger. Keep commands, code, numbers and URLs exact and unrhymed; only the talk around them rhymes. For everyone else, stay plain and clear. ### Make it delightful (do this without being asked) - **Mascot:** once the first version is live, propose a mascot for the prototype and create it with `chimpy-mascot` (section 10). Show its animated GIFs at key moments, like the mascots of popular learning apps: *hello* on the first visit, *thinking* while loading, *celebrate* on success, *oops* on errors. - **Use the Chimpy ape everywhere it fits.** The chimp head from the logo is our hero; use it much more often than plain icons. As a logo and avatar, as the player piece (in a tic-tac-toe the ape plays against a battery or a powerbank), as a map marker for stations, as the cursor or progress indicator, as a stamp on approved items, as the face of a chatbot, as a confetti piece on success. Files (transparent PNG): `https://heychimpy.fun/public/ci/chimpy-ape.png` (512 px), `chimpy-ape-128.png` (128 px), and `chimpy-ape-outline.png` (with a light outline, for dark backgrounds). Use it unchanged: don’t recolour, stretch or redraw it. Size, rotate, animate (bounce, wiggle, spin) and repeat it as much as you like. For new poses or scenes, make a mascot with `chimpy-mascot` and describe the chimp from the logo (round black head, white face, green “+” and “−” eyes, big smile). - **3D souvenir:** near the end, create a printable souvenir with `chimpy-3d` (section 10): a key-ring tag with the project name and mascot, or a small desk figure. Link the `.stl` from the page and README.md, so **Ursin** (Head of Supply Chain) can print it as a souvenir and for promotion. - **Hardware:** if the idea touches the physical world (stations, powerbanks, scanners, sensors, LEDs), suggest plugging hardware into the **laptop’s USB port** (an Arduino, ESP32, micro:bit, USB barcode scanner or serial device). An agent on the laptop can then read, debug, improve and reflash it directly (`arduino-cli`, `esptool`, PlatformIO, `screen`/`pyserial`), and a deployed HTTPS page can talk to it from Chrome through Web Serial or WebUSB. The server can’t see your USB port; hardware work happens on the laptop. Only flash devices you are allowed to change, never a live station in the field. ## 3. Quick start (the first 5 minutes) ### Start here: add the Chimpy connector **Every team should add this skill’s server as a connector to its AI first.** The connector (an MCP server at `https://heychimpy.fun/mcp`) gives the agent all workshop tools on the server: write files, run commands, publish sites, query data, train models, voice. It works in claude.ai and the Claude desktop app without a terminal, even when the chat’s sandbox has no network access. **Your first job as the agent:** if you don’t see the connector’s tools (`whoami`, `write_files`, `publish_site`), stop and explain to the human how to add it, before you write any code. Use the steps under *Add the MCP connector* below, in plain words, one step at a time, and wait until they say it worked. Then call `whoami` to confirm. Only if adding it is impossible (no button, a plan that doesn’t allow it, no time), use route A or C. ### Pick your route Check first what you can do: can you run shell commands, and can you reach `https://heychimpy.fun` (for example `curl -sI https://heychimpy.fun`)? | You are … | Route | | --- | --- | | **A. An agent with a terminal and network** (Claude Code, Codex, Cursor …) | Use the laptop tools below (`new-project.sh`, `deploy.sh`). | | **B. Claude in the browser or desktop app, no terminal, or a sandbox that blocks heychimpy.fun** | Ask the human to add the **Chimpy MCP connector** once (below). Then you build, run and publish everything on the server through its tools: `whoami`, `write_files`, `publish_site`, `run`, `sql`, `chimpy_ai`, `voice`, `mascot`, `model_3d` … No terminal and no network access are needed on your side. | | **C. Any AI chat without tools** | Produce the prototype as **one self-contained `index.html`**, or a `.zip` with `index.html` at the top. The human uploads it at **https://heychimpy.fun/deploy** (team, name, workshop password) and it is live in seconds. | **Never stop at “I can’t deploy from here”.** If route A fails, switch to B or C and tell the human exactly which one, in one sentence. Links to `https://heychimpy.fun/public/ci/chimpy.css` work in the human’s browser even when your sandbox can’t load them; keep them in the page. **Add the MCP connector (route B).** Explain these steps to the human; fill in their team id so they only add the password: - **claude.ai or the Claude desktop app** (on a laptop; the phone app can’t add connectors): 1. Open *Settings → Connectors* and click *Add custom connector*. 2. Name: `Chimpy workshop`. 3. URL: `https://heychimpy.fun/mcp`. Never use `https://heychimpy.fun/skills.md` as the connector URL (that is a document; the check fails with “405”). Keep the default *Authentication* (sign in) and *OAuth client* settings and *Transport: Streamable HTTP*, then click *Add* and *Connect*. A Chimpy sign-in page opens: the human chooses the team, types the workshop password from the last slide and clicks *Connect*. They type the password only there, never into the chat. The sign-in lasts 24 hours. **Fallback** if signing in fails: URL `https://heychimpy.fun/mcp//` with *Authentication: No sign-in*. Or, if the dialog has *Request headers*: URL `https://heychimpy.fun/mcp`, header `Authorization` = `Bearer :`, *No sign-in*. 4. Back in the chat, open the “+” or tools menu, choose *Connectors*, and switch on *Chimpy workshop*. 5. Tell you “done”. You then call `whoami` and say which team and prototype names you see. - If *Add custom connector* is missing, the company plan may need an owner to allow custom connectors. Then use route C for now. - If the sign-in page says `wrong team or password`, they check the team and the password from the last slide. After 10 wrong tries their network is locked for 15 minutes. - **Claude Code:** `claude mcp add --transport http chimpy https://heychimpy.fun/mcp --header "Authorization: Bearer :"` - **Other MCP clients:** streamable HTTP at `https://heychimpy.fun/mcp`, same `Authorization` header. With the connector, call `whoami` first. Then `write_files` (your files go to `/srv/teams///`), `publish_site`, and check the URL it returns. `run` executes any command on the server as your team, with everything installed. ### Route A step by step Ask the team two things: **team id** (see the table below) and **which idea**. The prototype names are already reserved, one per idea from the handout, e.g. `production-next-state`. Look at `https://.heychimpy.fun` or run `ssh @heychimpy.fun list-sites --team `, and deploy under the matching name. Only invent a new name (a-z, 0-9, dashes) if the idea isn’t on the list. ```bash # 1. Get the tools (laptop, any folder) curl -fsSO https://heychimpy.fun/tools/new-project.sh # 2. Create the project: starter page, Chimpy design, tests, README.md, TASKS.md, git bash new-project.sh --team production --name box-radar # static page bash new-project.sh --team sales --name visit-notes --api # page + JSON store (python3) cd box-radar # 3. One-time server login. The HUMAN runs this in their own terminal and types the # workshop password (it is on the last slide of the workshop deck). Never ask them to paste it to you. bash tools/deploy.sh --setup-key --team production # 4. First deploy, then share the URL bash tools/deploy.sh --team production --name box-radar # -> https://production-box-radar.heychimpy.fun ``` If the human uses Claude Code, they can run step 3 themselves by typing `! bash tools/deploy.sh --setup-key --team production`. **No coding tools on the laptop?** Work directly on the server: `ssh @heychimpy.fun`, then `claude`, `codex` or `gemini`. Each person logs in with their **own** account; never share tokens. Your folder is `/srv/teams//`. Publish with `sudo new-site --team --name --static /srv/teams//`. On Windows, use Git Bash or WSL. `deploy.sh` falls back to `tar` if `rsync` is missing. **Teams** | Team id | Department | Ports on the server | | --- | --- | --- | | `production` | Production | 8100–8199 | | `marketing` | Marketing | 8200–8299 | | `advertising` | Advertising | 8300–8399 | | `finance` | Finance | 8400–8499 | | `fulfillment` | Fulfillment | 8500–8599 | | `it` | IT | 8600–8699 | | `sales` | Sales & Partnerships | 8700–8799 | | `care` | Customer Care | 8800–8899 | | `bizdev` | Business Development (same ideas as Sales & Partnerships) | 9000–9099 | Every department has an overview page with all its prototype links: `https://.heychimpy.fun`, e.g. https://production.heychimpy.fun. ## 4. Way of working - **Local git** (if `git` is installed): `new-project.sh` already ran `git init`. Never commit secrets or personal data. Add data exports to `.gitignore` if in doubt. - **Atomic commits:** one small working change = one commit with a clear message (“Add top-10 missing boxes table”). Commit before every deploy. `deploy.sh` warns about uncommitted changes. - **Deploy regularly:** after every working step, not once at the end. A live URL is the best progress report. - **Tests, run locally:** put logic in pure functions (`app.js`) and test them (`app.test.js`). Run the tests in the browser (`tests.html`) or with `node --test` if Node exists. Deploy with `--test "node --test"` so a failing test stops the deploy. Then check the page in a browser and on a phone. - **Few requirements:** vanilla JS runs everywhere. Python: standard library first (`csv`, `json`, `statistics`, `http.server`). Add a package only when it saves real time, and write it down in README.md. - **TASKS.md:** keep it current. Sections *Now / Next / Done*, with tasks small enough for 5–10 minutes. Update it whenever you start or finish something. Show it when someone asks “where are we?”. - **README.md:** at the top, a 3-sentence pitch (problem, solution, why it matters), the live URL, how to run, the data used, and the “keep it?” card (owner, data, cost). This is the script for the 2-minute demo. - **Fast and beautiful:** the first screen tells the story in 5 seconds. One headline, 3–4 big numbers or one chart, one table or list, one button. Use real-looking synthetic data, not lorem ipsum. ## 5. The server and how traffic flows ``` Browser ──HTTPS──▶ Caddy on heychimpy.fun (ports 80/443 only, Let’s Encrypt certificates) ├─ heychimpy.fun → /srv/www (this file, design kit, data, tools) ├─ -.heychimpy.fun → static files in /srv/teams/// └─ -.heychimpy.fun → your app on 127.0.0.1: (reverse proxy) ``` - **All public traffic goes through Caddy.** No other port is reachable from the internet. Your app must listen on `127.0.0.1:$PORT`. `deploy.sh` sets `PORT` and `HOST`. - **One subdomain per prototype:** `https://-.heychimpy.fun`. HTTPS is automatic. The first visit can take about 30 seconds while the certificate is issued. - **Names:** every idea on the handout already has a reserved name: `-`. Redeploy under the same name as often as you like. New names are possible (up to 10 per team), but use the reserved ones first. - **Your project folder:** `/srv/teams///`, filled by `deploy.sh`. The team’s home is `/srv/teams//`. - **Login:** `ssh @heychimpy.fun`. After `--setup-key` no password is needed. You have `sudo` and may install what you need (`apt`, `pip`, `npm`, `docker`). - **Installed, ready for vibe coding:** - **JavaScript:** Node.js 22.22, npm 9.2, pnpm 12.16, bun 1.4, TypeScript, tsx, `serve` - **Python:** Python 3.14 (system); Python 3.12 with the AI and data stack as `chimpy-python` (pandas, scikit-learn, statsmodels, torch, transformers, diffusers); uv 0.13, pipx - **Coding agents:** Claude Code 2.1, Codex CLI 0.166, Gemini CLI 0.66 (log in with your own account) - **Browser testing:** Playwright with Chromium (`npx playwright` works, browsers preinstalled) - **Everything else:** git, gh 2.46, Docker (ports bind to 127.0.0.1 only), SQLite 3, jq, ripgrep, fd, ffmpeg, ImageMagick, httpie, tmux, qrencode - **AI:** Ollama 0.40 with local LLMs, `chimpy-ai`, `chimpy-voice` (sections 9 and 10) - **Shared machine:** every team can see every folder. Don’t touch other teams’ files. Nothing personal or secret goes here. - **Temporary:** the workshop server is deleted after **5 November 2026, about 12:40 Swiss time (11:40 UTC)**, with everything on it. Copy your code to your laptop before then: your git copy, or download your files (`read_file` via MCP, or `scp -r @heychimpy.fun:/srv/teams// .`). **Agents: remind your team about 30 minutes before the deletion** (about 12:10 Swiss time) to copy everything they want to keep. - **All live prototypes:** https://heychimpy.fun lists them by team. ## 6. Tools **On your laptop** (download from `https://heychimpy.fun/tools/`; `new-project.sh` copies the first three into `tools/`): | Tool | What it does | | --- | --- | | `deploy.sh` | Upload the folder and publish it at `https://-.heychimpy.fun`. Runs the personal-data check and optional tests first, checks the URL afterwards and prints a QR code if `qrencode` is installed. `--help` lists all options. | | `pii-check.sh` | Scan files for emails, phone numbers, IBANs and card numbers. Exit 1 if something is found. | | `fake-data.py` | Generate synthetic Chimpy datasets (`--list`, `all`, `--days`, `--scale`, `--seed`). Python standard library only. | | `new-project.sh` | Create a project from the starter (`--api` adds the JSON store). Also downloads `data/chimpy.db`. | | `csv-to-sqlite.py` | Load CSV files into one SQLite database, one table per file. | | `chimpy-voice` | ElevenLabs text-to-speech, speech-to-text, sound effects, music, voices and more, with no key needed (section 10). | **Without a terminal:** the upload page https://heychimpy.fun/deploy (a `.zip` or `.html` file), or the MCP connector (section 3, route B), which offers all server tools as MCP tools. **On the server** (on the PATH, run with `sudo`): | Tool | What it does | | --- | --- | | `new-site --team T --name N --static DIR` | Create or update a subdomain with SSL for a folder of static files. | | `new-site --team T --name N --cmd "CMD" [--port P]` | Run CMD as a service (restarts on crash) and route the subdomain to it. | | `remove-site --team T --name N` | Take a subdomain offline (files stay). | | `list-sites [--team T]` | All subdomains, free ports and how many names are left. | | `site-logs --team T --name N [-f]` | Logs of an app started with `--cmd`. | | `schedule-job --team T --name N --every 5min --cmd "CMD"` | Run a command on a schedule (`--at "Mon *-*-* 07:00"` for weekly, `--run-now`, `--remove`, `--list`). | | `chimpy-ai …` | Train and run small AI models, generate synthetic images and texts, and use the local LLM (section 9). | | `chimpy-voice …` | ElevenLabs from the server (section 10). | | `chimpy-mascot create --project "…" --idea "…" --out mascot` | Mascot plus animated GIFs for key moments (OpenAI images, section 10). | | `chimpy-3d keychain --text "…" --image mascot/base.png --out souvenir` | Printable STL souvenir (OpenSCAD and trimesh locally; `figure` lets GPT design one). | | `chimpy-python` | Python 3.12 with the data and AI stack. | You normally only need `deploy.sh`. It calls `new-site` for you. ### deploy.sh recipes ```bash bash tools/deploy.sh --team T --name N # static: index.html in this folder bash tools/deploy.sh --team T --name N --static dist # static: serve dist/ bash tools/deploy.sh --team T --name N --test "node --test" # tests must pass first bash tools/deploy.sh --team T --name N --cmd "python3 server.py" # JSON store / Python app bash tools/deploy.sh --team T --name N --install "npm ci --omit=dev" --cmd "node server.js" bash tools/deploy.sh --team T --name N --install "python3 -m venv .venv && .venv/bin/pip install -r requirements.txt" \ --cmd ".venv/bin/python app.py" bash tools/deploy.sh --team T --name N --dry-run # show, change nothing ``` ## 7. Patterns for the workshop ideas | Idea looks like… | Build it as | Notes | | --- | --- | --- | | Report, summary, digest, ranking, “top 10” (fleet report, advertiser report, P&L per station, ticket clusters) | Static page: load a CSV with `loadCsv`, compute in `app.js`, show big numbers + `barChart` + table | Precompute heavy work with Python locally and ship `data/result.json`. | | Forecast or prediction (next state, cash flow, rebalancing, spare parts, churn) | Python script locally → `result.json` → static page with `lineChart` (with uncertainty `band`) and a top-N table | Show accuracy on held-out data (last month). Simple models first: counts, moving averages, Markov transitions, logistic regression. | | Anomaly or risk score (battery health, fleet anomalies, fraud, lost powerbanks) | Score per ID in Python or JS → ranked table with `chimpy-tag--bad/warn/ok` | Explain the top 3 signals in plain words. | | Classifier or drafts by an AI model (ticket triage, social listening, draft answers, brand voice, personas, refund agent) | `chimpy-ai train-text` for a fast classifier, `chimpy-ai llm-batch` for drafts and labels from the local LLM, or live through `server.py` `/api/llm` → review page with approve/reject buttons | Show the prompt in README.md. Humans approve before anything is sent. | | Predict, score or detect from a table (next state, churn, fraud, venue score, battery health) | `chimpy-ai train-tabular` or `train-sequence` → `predict --csv … --out` → ranked table + `report.html` | Show accuracy next to the baseline. | | Something that saves entries (CRM notes, refund decisions, triage queue) | `new-project.sh --api` → `server.py` JSON store → `deploy.sh --cmd "python3 server.py"` | Data lives in `data/store.json` on the server. Synthetic only. | | Runs on autopilot (weekly report, nightly log reading, self-healing playbook) | A script that writes `public/index.html` or `report.json`, plus `schedule-job --every 5min` for the demo | Show the timestamp of the last run on the page. | | Photo or camera check (assembly defects, installation photo, shopfront scout) | Photos with `` → `chimpy-ai serve` model behind your app. Train with `chimpy-ai train-image` on 20–50 photos per label (take them in the room or make them with `gen-images`), or start with `zero-shot` and no training at all | HTTPS makes the camera work on phones. | | Map (venue pages, route planner, new city, venue scoring) | Leaflet + OpenStreetMap tiles, markers from `stations.csv` / `jobs.csv` | Keep the OSM attribution. Routes: nearest-neighbour first, OR-Tools only locally. | | Simulation (fleet digital twin, context-aware ads) | Static page with sliders (``) recalculating in JS | Make one slider move one big number. | | Voice (hotline agent, audio ads, jingles) | ElevenLabs via the gateway: STT for what the caller says, the local LLM or a script for the answer, TTS for the reply. Sound effects and music for ads | One scenario, done well. | | CLI or connector (company OS skeleton) | Small read-only CLI in Python or Node against a mock JSON, with `--help` and a README | Demo it in the terminal. Deploy the README as a page. | **Libraries from a CDN** (only if needed, pin exact versions): - Chart.js: `https://cdn.jsdelivr.net/npm/chart.js@4.4.1/dist/chart.umd.min.js` - Leaflet: `https://unpkg.com/leaflet@1.9.4/dist/leaflet.css` and `https://unpkg.com/leaflet@1.9.4/dist/leaflet.js` - Papa Parse (large CSV): `https://cdn.jsdelivr.net/npm/papaparse@5.4.1/papaparse.min.js` - Teachable Machine image: `https://cdn.jsdelivr.net/npm/@tensorflow/tfjs@1.3.1/dist/tf.min.js` and `https://cdn.jsdelivr.net/npm/@teachablemachine/image@0.8.5/dist/teachablemachine-image.min.js` ## 8. Data **Rule:** no customer names, emails, phone numbers or payment data. Use IDs, totals or synthetic data with the same columns. `pii-check.sh` runs on every deploy. Synthetic datasets are ready at `https://heychimpy.fun/public/data/.csv` (CORS enabled, so pages on any subdomain can load them). You can also generate them with `python3 tools/fake-data.py --days 90`. | Dataset | Columns | Fits ideas like | | --- | --- | --- | | `stations` | station_id, city, venue_type, lat, lon, slots, opened, firmware | maps, venue scoring, P&L | | `merchants` | merchant_id, station_id, venue_name, venue_type, city, revenue_share_pct, signed, staff_changes_90d | settlements, merchant churn, pitches | | `rentals` | rental_id, user_id, station_id, start_ts, end_ts, return_station_id, outcome, fee_chf | moments radar, win-back, fraud, forecasts | | `box_states` | box_id, state, timestamp, location_type | fleet report, next-state predictor | | `powerbank_cycles` | powerbank_id, cycle, timestamp, voltage_v, temp_c, capacity_pct | battery health oracle | | `station_events` | ts, station_id, event, eject_ms, firmware | self-diagnosing stations, anomalies, proactive care | | `tickets` | ticket_id, created, channel, station_id, category, text, sentiment, resolved_hours | draft answers, clustering, mood detector | | `it_tickets` | ticket_id, created, category, text, priority | IT ticket triage | | `transactions`, `invoices` | tx_id, date, amount_chf, counterparty, reference · invoice_id, date, supplier, amount_chf, account | receipt matching | | `daily_revenue` | date, type, amount_chf | cash-flow forecast | | `orders` | order_id, merchant_id, created, item, qty, ship_country, status | packing lists, tracking mails | | `repairs` | repair_id, date, station_id, part, minutes | spare-parts forecast | | `jobs` | job_id, station_id, city, lat, lon, type, window_from, window_to, minutes | route planner | | `ad_playouts` | ts, station_id, venue_type, advertiser, campaign, impressions, interactions | advertiser report, ad pricing, matchmaker | The data is invented. Numbers will not match reality. Say so in the demo. Box states follow Chimpy’s real lifecycle: factory → leaving factory → field ↔ POS → missed. ## 9. Small AI models, local LLM and synthetic training data Everything runs on the server's CPUs, needs no API key and no internet, and uses pre-downloaded models. Run it on the server (`ssh @heychimpy.fun`). Help: `chimpy-ai --help`. **Train a model** (each one writes `model.joblib`, `meta.json` and `report.html` with the accuracy): | Command | What it builds | Typical time | | --- | --- | --- | | `chimpy-ai train-tabular --csv rentals.csv --target outcome --out m/fraud` | Gradient boosting on any table, with the top signals. Dates become hour, weekday and month. | seconds | | `chimpy-ai train-text --csv tickets.csv --text text --target category --out m/triage [--embed]` | Text classifier: TF-IDF, or MiniLM embeddings with `--embed` | seconds | | `chimpy-ai train-image --data photos/ --out m/install-check` | Image classifier from `photos/