career.eval — a one-click triage for job ads.
The problem
Job hunting in Switzerland means reading 30 to 40 ads a week and triaging by gut feel. Each one needs cross-checking against three CV variants (ops, media, hybrid) to decide which version to send. Manual triage was eating 20 to 30 minutes per ad before the actual application even started.
The architecture
Single-page web app, runs entirely in the browser. Three tabs: paste a job description, store CV variants, store API key. One button. The Anthropic API key lives in localStorage, nothing leaves the browser except the API call itself.
Tools
AI: Anthropic API (Claude)
Storage: Browser localStorage
Outcome
Triage time per ad: under 60 seconds. The 20-minute manual cross-check became a one-click decision. Built and used in active job search.
What I'd do differently
v1 sends each CV as a fresh API call. v2 should batch them into a single structured prompt with comparative scoring, both for cost and consistency. Adding a "saved evaluations" history would help me track which framings actually got responses, and feed that back into the rubric.
A LinkedIn content program that ran 5 to 7x benchmark.
The problem
A B2B EdTech company with a quiet LinkedIn presence, no consistent publishing rhythm, and no analytics loop closing the gap between "what we posted" and "what worked." Marketing wanted a structured program. The team was small, the production cost of original video and design was high, and nobody had time to close the loop manually each week.
The architecture
A Notion calendar built around four pillars matching the buyer journey. Templates that let me produce a mix of static, motion, and short video without burning out the team. Claude for ideation and first-draft copy with human review and rewrite as the constant. A simple analytics tracker comparing reach, CTR, and engagement week-over-week. Weekly call on what to repeat, kill, or test next.
Tools
Production: Adobe Premiere Pro, After Effects, Illustrator
AI: Claude (ideation & first drafts)
Analytics: LinkedIn Analytics, manual weekly review
Outcome
63 posts over 12 months, two phases. 54,333 impressions. 9,231 clicks. Average CTR 15%. Average engagement 17%. B2B LinkedIn benchmarks sit at 2 to 3% CTR and around 2% engagement, so the program ran consistently 5 to 7x benchmark. 98% on-time publishing. Generated 10 inbound trainer applications and one direct sales lead attributed to LinkedIn visibility.
What I'd do differently
I'd close the analytics loop tighter, ideally automate the weekly digest into a Notion page so the "what worked" review takes 15 minutes instead of 90. I'd also build an experiment log earlier. We found the format-content combinations that worked, but I want them documented so the next person running the program doesn't have to rediscover them.
Studio Ops — a video pipeline that runs itself.
The problem
Solo content production has a hidden tax: the admin between the creative steps. Folder setup, file routing, metadata writing, status tracking. Roughly four hours per video. None of it the work I want to be doing. So I automated the boring parts.
The architecture
Three n8n workflows, all triggered by status changes in a Notion video pipeline database. Notion is the state machine.
Tools
State store: Notion (pipeline database)
File store: Google Drive
AI: Anthropic API (Claude)
Logic: JavaScript (parsing & folder rules)
Outcome
First episode through the full pipeline shipped end-to-end. Admin time per video: from ~4 hours to ~5 minutes. A 95% reduction on the operational layer. Studio Ops v2 in active build.
What I'd do differently
v1 polls Notion for status changes; v2 should switch to webhooks for lower latency. The metadata stage could also benefit from a manual review gate: the Claude output is good but not always what I'd publish first try. Cost tracking on Claude calls is missing, the metadata step is more expensive than I expected.