PRODUCTION PLANNING AGENT
29% of your factory
sits idle.
At 71% line utilisation your most expensive asset runs at three-quarter speed. Becci builds the optimal production schedule in 90 seconds and replans in 8 minutes after any disruption β recovering 23 percentage points of utilisation from the same physical plant.
Every capability below runs continuously β without anyone scheduling it, requesting it, or chasing it.
Becci does not plan against last week's data. Every decision is made against the current state of the factory floor.
Five steps. Becci handles all five. Your team receives the schedule, not the problem.
Documents and decisions produced autonomously β written directly into your MES, ERP, and inter-agent network.
A pre-recorded simulation of Becci replanning three production lines in 8 minutes after a hydraulic failure on Line B. Watch the full replan β then book a session to see it on your plant.
30 minutes. Manish runs Becci live using your actual plant data. You see your exact utilisation improvement, the replan in action, and the financial impact at your scale.
Configure your plant for responses specific to your production constraints and products.
3 production lines in real time.94%. Last schedule build: 90 seconds. Last replan event: none this shift.This paper presents Becci, the autonomous production planning agent within the VYUH Supply Chain Neural Network. Becci employs a Mixed-Integer Linear Programming (MILP) optimisation engine combined with a Travelling Salesman Problem (TSP) variant for changeover sequencing to generate and maintain production schedules across multi-line manufacturing environments. Becci achieves 94% line utilisation against a 71% industry-average baseline, builds complete optimal schedules in under 90 seconds, and replans following any disruption within 8 minutes. We describe the full technical architecture including MILP formulation, changeover sequencing algorithm, real-time MES integration, 8-minute disruption response architecture, confidence gating mechanism, and inter-agent coordination within the VYUH framework.
Production planning in multi-product, multi-line manufacturing environments presents a combinatorially complex optimisation problem. The planner must simultaneously satisfy constraints on machine capacity, raw material availability, labour, customer order priorities, changeover sequences, and maintenance windows β whilst maximising total output and protecting delivery commitments. The problem belongs to the class of NP-hard scheduling problems, meaning that even moderately sized instances cannot be solved to optimality by exhaustive enumeration in reasonable time.
Conventional production planning approaches address this complexity through simplification β planners apply rules of thumb, frozen horizons, and hierarchical decomposition to reduce the problem to a manageable scale. The result is sub-optimal schedules: industry average line utilisation of 68β72% against a theoretical achievable utilisation of 90%+. The gap represents enormous value β at NVIDIA's production scale, recovering 23 percentage points of utilisation from the same asset base is equivalent to adding more than one full production line without capital expenditure.
Becci closes this gap through autonomous, continuous schedule optimisation using a MILP engine that solves the full problem to provable optimality within the computational budget of a 90-second scheduling cycle.
Becci's architecture comprises five layers: (1) a real-time data ingestion layer connecting to MES, ERP, WMS, and inter-agent feeds; (2) a constraint formulation engine that translates live data into MILP problem parameters; (3) a MILP solver instance (Gurobi) with warm-start capabilities for rapid replanning; (4) a changeover sequencing module using a TSP-variant solver; (5) a confidence-gated publication and signalling layer.
The diagram below illustrates the improvement Becci delivers β the same three production lines before and after schedule optimisation, showing utilisation recovery from 71% to 94%.
Becci formulates the production scheduling problem as a Mixed-Integer Linear Programme over a rolling horizon of T time periods, M production lines, and N jobs. Binary decision variables x_{i,m,t} indicate whether job i is assigned to line m starting at time period t. The objective function maximises total weighted throughput subject to a set of hard constraints.
Becci uses the Gurobi solver with warm-start capabilities. For the weekly schedule build (full horizon), the solver runs for a maximum of 60 seconds, targeting a MIP gap of 0.5% β meaning the solution is guaranteed within 0.5% of the mathematical optimum. For emergency replanning, a reduced-scope formulation covering only affected jobs runs with a 90-second time limit and 1% MIP gap target. In practice, the 90-second replanning target is achieved in 7β8 minutes end-to-end including data ingestion and output publication.
Within the MILP solution, the sequence of jobs on each line is further optimised by a dedicated changeover sequencing module. The problem of finding the minimum changeover time sequence is formulated as an Asymmetric Travelling Salesman Problem (ATSP) β each product type is a "city" and changeover time between product types is the "travel cost". Becci solves this using the Concorde TSP solver for small instances and a Lin-Kernighan heuristic for larger portfolios, consistently recovering 8β12% additional output capacity versus naive (FCFS) sequencing.
The 8-minute disruption response is achieved through three architectural decisions. First, warm-start MILP: the solver is never cold-started β it maintains a feasible warm solution from the previous schedule and modifies only the affected portions. Second, scope reduction: the replan formulation covers only the jobs affected by the disruption, not the full horizon β reducing problem dimensionality by typically 60β80%. Third, pre-computed alternatives: Becci maintains a library of pre-computed replan templates for common disruption types (single line down, material shortage, quality hold) that can be deployed in under 2 minutes, with the MILP solving the exact optimisation in parallel.
Becci applies a confidence gate to each schedule publication. The confidence score is a composite of: MILP solution quality (MIP gap achieved), material availability certainty (whether all required materials are confirmed available), demand signal confidence from Sara Jr, and labour confirmation status. Schedules above 85% confidence are published directly to MES. Between 65% and 85%, a planner review flag is raised with a 2-hour default window. Below 65%, an escalation briefing is generated with alternative schedule options.
Becci receives demand signals from Sara Jr every 6 hours, material availability confirmations from Cho in real time, and procurement delivery confirmations from Ari. Becci publishes material pull-forward requirements to Cho and finished goods despatch timing to Tom. This bidirectional integration means a demand uplift signal from Sara Jr is reflected in the production schedule within the next scheduling cycle β typically within 6 hours β without any manual planning intervention.
Becci's MILP approach has documented limitations in three cases: highly stochastic environments where yield rates vary significantly between runs (addressed by safety time buffers in the formulation); very large portfolios exceeding 500 jobs per planning horizon (addressed by hierarchical decomposition); and novel product introductions without established processing time data (handled through conservative time estimates and escalation flagging).