Shift Sight Ops

An operations copilot for shift leads: a WMS tells you what happened — the copilot tells you what is going to happen, and what to do. The problem. A shift lead in a fulfillment building finds out the night is lost when it is already lost. Warehouse systems report what already happened; by the time a staffing shortage or a stopped line shows up in the numbers, the cutoffs that will miss are already decided. The lead's real job — seeing the night before it happens and moving people in time — is done by instinct. What Shift Sight Ops does. The copilot projects the night's on-time shipping (OTS) minute by minute from the current state of the floor: staffing by area, the volume plan, the cutoff calendar. Before the shift starts it can already say whether the plan as staffed will make the night. During the shift it detects line stoppages from the aggregate scan-gap signal (an area with no scans for 5 minutes), recomputes the projection, and recommends staffing moves in the floor's donor order, each one chosen for the OTS it recovers — rebalancing people across Picking, Induct, Put, Pack and Shipping, at a realistic 10-minute relocation cost per move. When rebalancing is not enough, it calculates the remaining shortfall in units and drafts the extra-staff posting — area, headcount, hours — for the lead to approve. It never sends anything. How it works — three actors, three channels. A deterministic engine watches and recommends on its own: a pure-Python simulation engine (no dependencies) computes every number — capacities, per-cutoff OTS projections, stoppage detection, donor recommendations, posting shortfalls. It is the only source of figures in the system. The language model only speaks when the lead asks: when the lead asks a question, it calls the engine as a tool, reads the result, and explains it; the model is instructed never to state a figure that did not come from a tool result, and in the demo every figure it states traces to an engine call. If the model API is unavailable, the copilot degrades to the engine's own templated explanations: the numbers and the alerts never depend on the model. The person always decides: recommendations and drafted postings are proposals; nothing moves and nothing is posted until the lead says so. The demo night (all data synthetic). One night shift, 17:30–03:30, a 100,000-unit day, 12 cutoff times, a five-node chain (Picking → Induct → Put → Pack → Shipping), two line stoppages, and absences that hit the usual donor area itself — so the night cannot be saved by rebalancing alone. Before the shift, with the roster as planned, the projection is 94.0% OTS against a 99.0% target — the night is at risk before a single box has moved. The copilot recommends five staffing moves; with them applied, the projection rises to 99.3% and no posting is needed. At 20:00 the Picking → Induct line stops and the live projection drops to 73.1%. At 20:15 the no-scan signal confirms the stoppage; the copilot recommends four more moves, lifting the projection to 76.2%, and drafts a posting (5 people to Picking, 5 hours; shortfall ~2,105 items). The posting is drafted, not simulated — the simulator only moves people already on the roster. Played out to the end of the night in the full simulation, the same night closes at 69.5% realized OTS with no intervention, 73.1% with only the pre-shift moves, and 76.2% with both sets of decisions. One honesty note on the stoppages: in the simulation, a stoppage enters the projection the moment it starts; on a real floor it would enter only through the 5-minute aggregate no-scan signal. Live demo board: https://albertojbravo-art.github.io/Shift-Sight-Ops/web/ — playable in the browser (static replay of the engine traces and the copilot's real exchanges; no live API calls). Team: Alberto Bravo — solo developer (Indie Dev track). I designed and operate the project, and I built it with the assistance of my AI agent Muse (Meta's personal AI agent, which I call Orion); the repository's construction commits are authored by that agent's account (orion-collab28). The product's copilot itself runs on Meta's Muse Spark Open. Data: All data in this project is synthetic. No real employer data, no real people and no real operations figures are used anywhere in the code, the scenario, the dashboard or the demo video. The demo night is a hand-checkable synthetic scenario, documented in the repository (SCENARIO.md). Transparency: Character and voice AI-generated; voice cloned from Alberto, with his consent.