Not slides. Real systems, drawn from the actual workflow files and the actual code. Open any of them for an interactive canvas you can pan and zoom, and a plain explanation of how it works, what it is for, and what broke along the way.
Scores every production voice call for quality and, separately, for compliance. Four layers, cheapest first.
Listens for call_end, judges six dimensions with strict evidence rules, posts to Slack only when a human is needed.
ASR to LLM to TTS on LiveKit, browser and phone from one worker. 902 ms end to end, measured.
Scores every workflow on an instance against nine fragility rules. HTML report on demand, Slack summary weekly.
Groups audit scores by account, computes composite health, routes critical to an intervention and leaves healthy alone.
Searches 18 titles daily, knocks out misses with rules, scores the rest with GPT-4.1, writes to Notion.
Classifies inbound mail, matches it to the application, writes status back, drafts a reply to rejections. Never sends.
Ranks contacts by fit, spends enrichment credits only on the right ones, drafts a German letter into each row.
One small workflow per tool: verify the caller, find care, book or move an appointment, answer in a shape the agent can say.
Onboarding, a six-signal discovery feed and behavioural notifications designed as one activation loop.
Berlin-based, working across the EU and the Gulf. Happy to walk through any of these on a call.