Cho manages inventory across every warehouse, every SKU, every echelon — simultaneously. Multi-echelon optimisation calculates the precise safety stock needed at every node, frees trapped working capital without ever risking a stockout, and reorders autonomously with 99.97% accuracy.
Cho eliminates both stockouts and excess. Not one or the other — both.
Cho never relies on a single number. It cross-validates every signal across eight independent streams.
A pre-recorded simulation of Cho analysing 12 warehouses across 2,400 SKUs and identifying £180M of working capital locked in excess safety stock — then generating the phased release plan. This is what your inventory would look like.
Configure your operation below and every response becomes specific to your warehouse network, SKU base, and supply chain.
This paper presents Cho, the autonomous inventory management agent within the VYUH Supply Chain Neural Network. Cho employs a stochastic multi-echelon inventory optimisation model to calculate globally optimal safety stock levels and reorder points across complex distribution networks. In deployment across a 12-DC, 2,400-SKU electronics distribution network, Cho has reduced average safety stock from 88 days to 22 days, released £600M in working capital, maintained 99.97% inventory accuracy, and achieved 91% autonomous replenishment execution. We describe the full technical architecture, mathematical formulation, confidence gating mechanism, and inter-agent coordination framework.
Inventory management in complex distribution networks presents a fundamental tension between service level and working capital efficiency. Traditional approaches — whether spreadsheet-based reorder point models or first-generation ERP replenishment modules — address this tension through static safety stock calculations, periodic review cycles, and conservative buffer policies. The result is predictable: excess inventory accumulates to protect against uncertainty, working capital is trapped at scale, and stockouts still occur because the static models cannot respond to real-time demand and supply signals.
The typical electronics distribution company holds 60–90 days of safety stock across its network. Academically optimal levels — calculated from actual demand variability and supplier lead time distributions — are typically 18–25 days. The gap between these numbers represents a substantial fraction of a company's working capital, earning no return and degrading the balance sheet.
Cho closes this gap through autonomous, continuous multi-echelon optimisation — recalculating optimal inventory levels daily using live signals from across the VYUH agent network and executing replenishment decisions autonomously for all routine cases.
Cho's architecture comprises four integrated layers: (1) a real-time data ingestion layer connecting to WMS, RFID sensors, GS1 scan systems, and inter-agent feeds from Sara Jr and Ari; (2) a multi-echelon optimisation engine that formulates and solves the network-wide inventory allocation problem daily; (3) a real-time monitoring and triggering layer that continuously compares live inventory positions to dynamic reorder points and queues reorder signals; (4) a confidence-gated publication layer that routes autonomous signals to Ari and escalates exceptions for human review.
The system maintains a live digital twin of the entire inventory network — every SKU, every location, every in-transit quantity, every pending reorder. This twin is updated on every WMS transaction, every Ari PO confirmation, and every Sara Jr forecast cycle — ensuring the optimisation engine always operates on current data.
Cho formulates the inventory allocation problem as a stochastic multi-echelon optimisation. The network is modelled as a directed graph where nodes represent inventory holding locations (central DC, regional DCs, forward positions) and edges represent replenishment flows with associated lead time distributions.
For each SKU i at location j, Cho calculates the optimal base stock level S*ᵢⱼ that minimises the expected total cost:
Minimise Σᵢⱼ [hᵢⱼ · E[Iᵢⱼ] + bᵢⱼ · E[Bᵢⱼ]]
Subject to: service level constraints per SKU category, network flow conservation, capacity constraints per location, and budget constraints on total inventory investment. Here hᵢⱼ is the holding cost per unit-day, bᵢⱼ is the stockout penalty, E[Iᵢⱼ] is the expected on-hand inventory, and E[Bᵢⱼ] is the expected backorders.
Demand at each node is modelled as a compound Poisson process with parameters estimated from the most recent 24 months of historical data and updated with Sara Jr's real-time demand signals. Lead time distributions are maintained per supplier per SKU from Ari's confirmed delivery history, updated on every completed purchase order.
The safety stock at any node is derived from the base stock level minus the expected demand during lead time: SS = S* - μL, where μL is the expected demand during the lead time. The key innovation relative to traditional methods is that both S* and μL are recalculated daily — incorporating current demand variability from Sara Jr and current lead time distributions from Ari — rather than set once from historical averages.
This dynamic recalculation is the mechanism through which Cho reduces safety stock without increasing stockout risk. As Sara Jr improves forecast accuracy over time, the demand variability estimate decreases — allowing safety stock to shrink safely. As Ari negotiates more reliable supplier lead times, the lead time variance decreases — allowing further safety stock reduction. The 22-day figure is not a target; it is the current mathematical optimum given current forecast accuracy and current supplier reliability.
Cho applies a composite confidence score to each reorder signal before execution. The confidence score integrates: forecast confidence from Sara Jr (how certain is the demand signal driving this reorder?), supplier lead time reliability from Ari (how consistent is this supplier's lead time?), inventory data freshness (how recently was this SKU's count physically verified?), and reorder quantity normalcy (is this quantity within the normal range for this SKU, or is it anomalously large?).
Signals above 85% confidence execute autonomously to Ari. This threshold is calibrated to deliver 91% autonomous execution — meaning the team handles approximately 9% of reorders that involve some uncertainty. 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 reorder quantities and context.
Cho operates as a tight node within the VYUH agent network. Sara Jr's demand forecasts arrive every 6 hours and update Cho's demand models — immediately propagating any forecast revision into the inventory position projections and reorder point calculations. Ari's purchase order confirmations update Cho's in-transit inventory position in real time, preventing duplicate reorders. Tom receives inter-warehouse transfer instructions from Cho for rebalancing movements, coordinating the freight logistics autonomously.
This inter-agent architecture eliminates the information latency that causes most inventory management failures. In traditional environments, a demand forecast change takes days to propagate into a revised reorder decision. In the VYUH network, it propagates in minutes — before the inventory position has time to deteriorate.
In 90-day deployment at a 12-DC, 2,400-SKU electronics distribution network: safety stock reduced from 88 days average to 22 days (75% reduction). Working capital released: £600M. Inventory accuracy improved from 95.3% to 99.97%. Autonomous replenishment rate: 91%. Stockout frequency reduced by 75%. Carrying cost reduction: £12.6M per annum.
The working capital release follows a characteristic phased pattern: the first 30 days deliver the largest single tranche as the most obviously over-stocked SKUs are identified and release plans initiated. Months 2–4 deliver progressively refined optimisation as the demand models improve with additional data.
The current implementation has documented limitations in three areas: highly promotional environments where demand exhibits extreme short-term spikes that the Poisson model underestimates; networks with extremely long tail SKUs (99th percentile demand velocity very low) where sparse data limits the accuracy of demand variability estimates; and novel product introductions where there is insufficient historical data to calibrate lead time distributions. Each limitation is handled through conservative fallback parameters and escalation flagging.
Future development priorities include: integration of external demand sensing data (social media, competitor stockouts) into the demand model; dynamic service level targets that adjust by SKU profitability; and a multi-period optimisation formulation that jointly optimises ordering costs with holding costs across the full planning horizon.