Inventory Management

The core challenge: multiple channels placing orders concurrently against the same stock. Race condition can lead to oversell. Solution: 1) Unified inventory pool with channel buffers: Each channel gets a 'soft allocation' (e.g., Amazon channel buffer = 100 units of SKU X). Order above buffer requires a real-time check against shared pool. 2) Redis for real-time reservation: On order receipt, DECR command on Redis counter (atomic). If result ≥ 0, reservation succeeds. If < 0, increment back and reject. Redis ensures atomic, sub-millisecond reservation at high concurrency. 3) Order state machine: RESERVED → CONFIRMED → FULFILLED → SHIPPED. Each state change emits a Kafka event. 4) Reconciliation job: Every hour, compare Redis (reserved) vs DB (confirmed orders) — fix any drift. 5) Channel stock push: After each reservation, recalculate available qty per channel and push to marketplace via API. Delay = ~5 min. Buffer prevents oversell during the delay window. Scale: Flipkart handles 5M+ orders/day = ~58 orders/sec average. Peak = 10x. Redis easily handles 100k ops/sec.

FIFO (First In, First Out): Oldest stock consumed first. During inflation, FIFO gives lower COGS (old cheaper stock used) → higher profit → higher taxes. Inventory on balance sheet = recent (higher) prices. Used in most perishable goods. LIFO (Last In, First Out): Newest stock consumed first. Higher COGS during inflation → lower profit → tax advantage. Not permitted under IFRS (India follows Ind AS = IFRS-based) — only allowed under US GAAP. WACC (Weighted Average Cost): Each stock movement uses an updated average cost = Total inventory value / Total quantity. Simpler to maintain. Smooths out price fluctuations. Most common in India for non-perishable goods under Ind AS. System implications: WMS/ERP must track lot/batch costs for FIFO. WACC is simpler — just maintain unit cost and update on every receipt. Pharma and food typically FEFO (First Expired, First Out) — variant of FIFO but based on expiry date, not receipt date.

Forecasting process: 1) Baseline statistical model: Time-series models on historical sales — ARIMA, ETS (Exponential Smoothing), or ML (LightGBM). 2) Causal adjustments: Add promotions, events, seasonality, price changes as features. 3) New product: Attribute-based similarity forecasting — find similar existing products and map their lifecycle. 4) Consensus: Final forecast = statistical baseline + planner override. Key accuracy metrics: MAPE (Mean Absolute Percentage Error): |Actual - Forecast| / Actual × 100%. Simple and intuitive but undefined when actual = 0. WAPE (Weighted Absolute Percentage Error): Sum(|Actuals - Forecasts|) / Sum(Actuals). Handles zero-demand items better. Bias: Avg(Forecast - Actual). Positive bias = consistently over-forecasting → excess stock. Negative bias = under-forecasting → stockouts. MASE (Mean Absolute Scaled Error): Scales error against naive forecast. Good for comparing across SKUs. World-class MAPE for FMCG: ~15–20%. For fashion (long lead time, high volatility): ~30–40% still acceptable. Key insight: Forecasting accuracy decreases at lower aggregation levels (store × SKU level harder to forecast than national level).

Single-location safety stock: Each node (DC, store) independently calculates its own safety stock based on its own demand variability and supplier lead time. Problem: Each node buffers against uncertainty independently → total system inventory = sum of all buffers → massive over-stocking. Bullwhip effect: Small demand fluctuations amplify as you go upstream in the supply chain — store sees +10% demand spike → DC buffers +20% → factory plans +40%. Multi-echelon inventory optimisation: Models the entire supply chain as a system. Optimises total system inventory while meeting service level targets. Key insight: Inventory held upstream (DC) can serve multiple downstream nodes (stores) → pooling effect reduces total required inventory. Safety stock formula accounts for both demand variability at each node and the uncertainty of replenishment from upstream. Tools: Blue Yonder, SAP IBP, Llamasoft (now Coupa). In practice: Amazon, Walmart use sophisticated multi-echelon models. A 3-echelon model (supplier → DC → store) can reduce total inventory 20–30% vs single-echelon while maintaining the same service level.

Constraints: 2,000–3,000 SKUs in a 2,000 sq ft dark store; 10-min delivery = pick time < 2 min; high perishable content. System design: 1) Store layout: ABC slotting — top 20% SKUs (A-class) at front bins (fastest to pick). Perishables in temp-controlled zones. Smart bin assignment minimises picker travel. 2) Real-time stock: Every item tracked at bin level with RFID or barcode. Stock updates on every order pick. Zero tolerance for phantom inventory — causes delivery failure. 3) Hyperlocal demand forecasting: ML model per store × SKU × hour-of-day. Monday 8pm ≠ Tuesday 8pm ≠ Sunday 8pm. Training data: past orders + weather + events (IPL match → beer demand spike). 4) Replenishment from mother warehouse: When bin stock < safety stock (e.g., 2 units), auto-trigger replenishment. Mother warehouse dispatches to dark store on intraday cycle (every 3–4 hours). 5) Availability on app: When stock = 0, item shows 'Out of Stock' immediately. This is critical — better to show OOS than fail delivery. 6) Expiry management: FEFO picking logic. Items expiring today flagged for markdown / donation. 7) Wastage tracking: Perishables wasted recorded for P&L and forecast model feedback.