Live work-skills calibration with merit-based XP

A live calibration system that reads how each user performs across five work-skills (independence, quality, communication, speed, and an overall roll-up), updates that read continuously as they work, and turns it into experience points that match the user's economic level on the platform to where they would actually stand against real job requirements. The substrate is the single shared layer the rest of the platform reads from when it needs to know who the user currently is.

Most learning platforms grade tasks but never grade the worker. To match a user to real job opportunities, the platform needs a live read of the qualities hiring managers actually look at: how independently they push through unknowns, how well their work holds up, how clearly they communicate, how fast they move. And that read has to keep updating as they work, not be a one-time snapshot from a final exam. Without it, XP is just a vanity number. With it, XP becomes a defensible signal of where the user currently stands relative to a real role.

Every touchpoint (chat turn, note submission, voice session, work submission, scenario completion, presentation) fires its raw signal into the substrate. A Temporal workflow runs as a long-lived per-learner singleton (one workflow instance per user, persistent across sessions) and routes every event through two paths in sequence. The cheap path is deterministic and runs on every micro-event: it reads the user's current work-skills snapshot, picks an XP base for that touchpoint, multiplies by the tier ladder from the overall work-skill (the multiplier ranges from 1x for underperforming up to 10x for exceptional), and writes the XP into an append-only event ledger plus a denormalized total. The heavy path fires when a touchpoint completes (stage complete, presentation complete, kickoff understanding complete, note-checker attempt complete, cowork session end). It runs an LLM chain that reads the full transcript, the rubric, the user's prior calibration entries, the cowork session events, the learn-mode usage, and the current snapshot, then rewrites the work-skills snapshot in place with a fresh per-skill verdict. The next cheap-path event reads back the freshly rewritten snapshot, so XP awarded on the very next interaction already reflects the latest verdict. The same substrate is what phase calibration, behavioral state, and the journey map all write into, so the rest of the platform reads one shared layer instead of pulling from five different places.

Tech lead. Owned the architectural design of the substrate, the cheap deterministic path that runs on every micro-event, the heavy LLM path that fires at touchpoint completion, and the XP allocation that ties tier to merit. Same substrate is what phase calibration, behavioral state, and the journey map all write into.