Parts list: projects

Arkon Manufacturing AI

An Industry 4.0 quality platform for a fictional heavy manufacturer, deployed on a NAS and running. Seven machine-learning models on real public datasets (remaining useful life of test engines, fleet fault classification, four kinds of visual inspection, a text model over consumer complaints) publish one twelve-field risk event, so nothing downstream knows which model spoke. A Quality Steering Cell on n8n validates the contract, suppresses repeats, writes an append-only incident store, assigns by department and puts a Telegram card in front of a named person for a P1 or P2. A Streamlit cockpit is the only surface that moves an incident; a Langflow assistant answers why, grounded in ten documents and the live store, and hands the operator a filled form instead of signing it, so the response-time KPI still measures the plant and not the agent. Two agents face opposite ways: the Langflow assistant inward, to the plant, and an n8n Customer Quality Desk outward, to the customer, which answers an OEM asking after a quality notice by its reference. They meet in exactly one file and the traffic runs one way: eight customer-safe fields out for one reference, no write endpoint back, and a guardrail in front of the desk and a sanitizer behind it. A live emitter raises a real re-timed incident every eight to twelve minutes and a simulated crew works it. The executive view is a Tableau workbook generated as XML from the same status API and published on Tableau Public.

Both agents answered as live demos on the internet until 15 September 2026, behind a login and on a temporary OpenRouter key that ends in mid-September 2026. Both public routes are closed now; the repository README shows screenshots of each agent at work, and the two Tableau views need no key at all. The plant, its crew and the incident stream are simulated and labelled so; the models, the datasets, every timestamp and the n8n, Langflow, Streamlit and Tableau layers are real.

AI-Brain Agent System

A job search, a full-time data programme and a set of personal admin domains all generate the same shape of recurring work: gather from a source, deduplicate against history, judge each item, write the result to the right file, and lose nothing between sessions. Doing it by hand does not scale, and handing it wholesale to a chat assistant fails differently: an assistant has no memory between sessions, will redo work that is already done, and will produce a plausible answer rather than report that it could not get one. So the design question was never "can an agent do this" but "how do you stop it from silently doing the wrong thing". The answer is four rules that everything else follows from. Every recurring job names exactly one scheduled owner and one canonical output path, because several agents write the same files with no database lock. A deterministic Python layer does the fetching, parsing, deduplication and writing, and the model is handed only the judgement that actually needs judgement. Verification gates can fail a run: separate checks assert that every identifier claimed in a report exists in the registry and the archive, that no encoding corruption shipped, that generated PDFs are machine-readable in the right order, that no document claims a skill the source-of-truth file does not carry, and that the prose does not read like a template. And the watchdog runs on a machine that cannot fail together with the machine it watches, because a watchdog sharing fate with its subject reports nothing exactly when it matters. The failure modes are the real result and each one was found by running the thing: fabricated metrics offered as interview material, a drafting agent absorbing the source document's vocabulary into a description of the candidate, and the class no gate catches at all, where a documented method is escalated by the verb rather than the noun. That last one is written down as a stated limit, not as a solved problem.

No public repository: the system operates a private vault. Designed, specified and operated by me; the code was written largely with AI assistance. Demonstrable live by screen share.

FreshMart Demand Planner

End-to-end retail demand forecasting: EDA, feature engineering, 10+ models compared (SARIMAX, Prophet, XGBoost, LSTM), HyperOpt tuning, and MLflow experiment tracking. Deployed as a Streamlit planning tool that speaks the language of store managers: total demand, peak day, average daily forecast.

Predictive Maintenance: Fleet Risk & Cost

Three business questions answered on a 92,000-row logistics fleet dataset: which vehicles need maintenance (classification), what it will cost (regression), and how to segment the fleet by risk (clustering). SHAP analysis identified brake condition as the dominant failure predictor, exactly what 17 years in vehicle engineering would suggest.

AI Order Triage Agent

A 26-node n8n workflow that triages free-text wholesale orders in English and Portuguese. Deterministic Airtable lookups gather every fact first; one bounded LLM agent makes the single call a rule cannot express; logic routes and logs every order, and a human stays the sender: the workflow never contacts a customer. Across three runs of each of the ten sample orders, all 30 recorded classifications matched the reference answer key, and no input ever produced two different classifications. Built on a fictional wholesale case, Volta Coffee Roasters in Porto.

Retail KPI Dashboard

Interactive Tableau dashboard tracking sales, profit, discount, and orders with drill-downs by region, product category, and time. LOD expressions and dashboard actions give non-technical users self-service analysis, the same logic as a quality trend board in a plant.