An AI-powered skincare recommendation system designed around three increasingly proactive modes: Assurance, Advisor and Anticipator.

Assurance evaluates product-customer fit on the product detail page, providing match information with rationale and anchor-based alternatives for mismatches. Advisor sits on customer profiles, using diagnostic inputs to rank the top five products across seven categories. Anticipator proactively predicts skincare needs to deliver timely recommendations rather than relying on explicit intent.

Skincare recommendation quality is a trust problem requiring suitability, explainability, and evidence.

Suitability spans skin type, concerns, product traits, ingredients, and reviews amid inconsistent catalogue data. We built a structured data layer to bridge gaps and combined retrieval-ranking logic to prioritize the most relevant options, complete with evaluations for false-positive risks.

Zalando Assistant answers shopping questions in natural language. It was built for fashion. I took it into beauty — a category where "what should I use?" or "does this product fit me?" carries consequence and health risk.

That meant retrieval over ingredient, suitability and review data the fashion index never held; a routing decision on when the assistant should recommend, ask, or decline; and an eval set of real journeys scored before launch.

The failure mode in a general assistant isn't a wrong product. It's a confident answer to a question it should have refused. I mapped where the existing architecture broke under beauty queries, then set the boundary: what it grounds, what it hands off, what it stays quiet on.

Doing that meant learning someone else's system well enough to argue with it — retrieval path, eval harness, failure points — and landing changes through a team that wasn't mine.

The page now opens with AI-generated product attributes — skin benefits, sensitivity, the top notes of a perfume — produced at a catalogue scale manual content never reaches.

Around it: curated storytelling for luxury brands, a visual 40+ shade selector in place of a size dropdown, and a How to Apply section the page never had.

374M visits a year, converting at 2.7% against a 3.4% benchmark. The gap was structural: a page built for fashion, used to evaluate beauty. Fixing it meant moving product data, content production, brand, legal and three engineering teams together, on a live surface, and cutting 26 candidate initiatives down to four milestones.

Rebuilt the flash sale engine with machine learning to automate product selection, discount depth, and slot timing across 1M+ SKUs for 300M+ customers.

To replace the manual system, I introduced five structural changes: switching to a unified product feed, ranking placement by discount depth, assigning single-point ownership with explicit KPI targets, digitizing pre-planned monthly budgets, and allocating spend toward true customer demand over high-ticket items.

Flash sales require multi-objective optimization — balancing margin, sell-through, and traffic within tight timeframes across a massive catalogue. Manual curation could not manage all three parameters simultaneously, requiring us to solve complex algorithmic trade-offs while building merchandising team trust in automated decisions.

Hastings is a personal agent that runs on my own machine, built on OpenClaw, an open-source gateway that keeps it local and wired to Telegram. It wakes at six, reads my personal Gmail, my LinkedIn inbox, a fare watch on Berlin to Mumbai and about forty news sources, and has one brief ready by 06:01.

Every card carries buttons. Send the reply to the broker in Kochi. Post the LinkedIn response. Arm the fare watch at €470. File the AI items into my wiki, cross-linked to the essay draft they contradict.

Ranking, not retrieval. Deciding which twelve items deserve six minutes of my morning needs a view of what I am working on this week, so the wiki feeds back into the ranker instead of only receiving from it.

Drafting in my register. The first version wrote replies I would never send. What fixed it was my sent mail and my own writing rules, then cutting the instructions to a tenth of their length.

Fares are a search problem. Hastings keeps a price history per route and interrupts me only when the curve breaks its own trend.

Blast radius. Personal accounts only, work mail never connected, nothing sent without a tap, every action in a ledger I can read. My Zalando context sits in a separate setup Hastings has no route to. A capable agent with the wrong permissions is a worse outcome than a dumb one.

Widening the scanner past RSS into X and YouTube, where most AI news breaks first. Then price comparison across retailers, a harder ranking problem and closer to my day job.

Vantage is a small web app I built and deployed myself. It's a chat coach for tech professionals working through a moment where how they come across matters — a promotion interview, a panel, a meeting with senior leadership in an unfamiliar business culture. You tell it what's happening, answer a few quick questions, and it gives you one concrete move for the conversation plus an outfit recommendation tied to your actual situation.

The interesting part is what happens after you close the tab. Sign in again next week and Vantage opens with your name and a reference to what you last talked through, instead of starting over.

Getting the model to output clean, structured data. Vantage pulls facts out of a normal conversation — your role, your name — and writes them to a database. Asking the model for JSON and nothing else sounds simple, but it kept wrapping the JSON in a markdown code block anyway, out of habit. Every extraction silently failed for a while before I noticed, because I'd built it to fail quietly rather than break the conversation. The fix wasn't a better prompt. It was code that strips the wrapper before parsing.

Deciding what the model is allowed to be confident about. The outfit recommendations were the riskiest part. It would have been easy to have the model suggest a specific product or price, all invented, since there's no live catalogue behind this. The rule I landed on: real, well-known brands as a starting point, never a SKU, a price, or an availability claim. Getting a language model to be useful without quietly overstating what it knows took more work than the feature itself.

Other use cases for generating executive presence — communication tips, networking topics, meeting tips. A real profile and memory system, recognising when someone's role has actually changed instead of just letting the old answer expire. And a real product catalogue, so the outfit advice can go from "try Uniqlo or COS" to something it can actually point you to.