WorkLens (PurpleAI Labs)

WorkLens turns a screen recording into a process map. Most business processes never produce an event log: they run across email, PDFs and several web apps, and the only record is the screen. A Chrome extension records a task with the person's role and the process it belongs to; a worker samples the video, finds step boundaries by pixel-diff, uses SAM 3.1 to segment exactly what changed at each boundary, and has Muse Spark name every step from the frames (application, action, object, fields seen). A deterministic graph and rule set then find loops, waits, re-keying between apps and long steps, each with the frames as evidence. Several recordings, one per role, merge into one process: handoffs are matched on the identifiers both screens showed (invoice and PO numbers), so WorkLens sees what no single recording can - a check done twice by two roles, an approval that leaves the system of record to travel by email. Recommendations come in two paths. Fix the process: rules turn each finding into a change, with the saving computed from the measured durations and the method shown. Apply AI: Muse Spark classifies each step against a fixed menu of seven mechanisms (document extraction, lookup and validation, pre-fill, drafting, in-app approval, notification, straight-through) and states what stays human; decisions are never automated. Accepting a recommendation updates the target process (handling time, steps, app switches, handoffs) and is recorded with who and when. Everything exports as JSON and BPMN 2.0. Demo: a real invoice approval across Gmail and Odoo, two roles, recorded with the extension. Live app: https://worklens-iota.vercel.app Stack: Meta Model API (SAM 3.1, Muse Spark), Python worker on Modal, Supabase, Next.js on Vercel, Chrome extension (MV3), ffmpeg. 45 tests on the rules and savings maths. Savings are labelled estimates.