In our Engineering Intelligence series, we highlight the architects behind the breakthroughs in supply chain autonomy. Today, we spotlight Parth Bisht , AI Engineer Lead. His team is moving beyond the reactive alert culture to build a system that doesn’t just predict problems, but fixes them.
Explore how the SpectraONE team:
- Eliminated decision latency by building a decoupled optimization layer.
- Transformed “data exhaust” from fragmented ERPs into high-fidelity signals.
- Solved the AI “black box” problem with mathematically justified decision paths
The Shift from Prediction to Decision Optimization
What architectural shift was required to move from reactive “prediction pipelines” to prescriptive “action engines”?
From prediction pipelines to decision optimization
Legacy supply chain AI is obsessed with state estimation telling you a shipment will be three days late. But for an engineer, a prediction isn’t an output; it’s just a high-uncertainty input. We designed our engine for decision optimization.
Embedding a reasoning layer for real-time decisions
Architecturally, we moved away from linear pipelines. Instead, we embedded a reasoning layer comprised of constraint solvers and agentic planners directly on top of our predictive models. We treat the prediction as one variable in a decision graph that evaluates thousands of “what-if” scenarios in real-time. We shifted the engineering focus from pure model accuracy to decision quality under constraints (cost, SLA, and capacity)
Decoupling Intelligence from the ERP Monolith
How does the engine return actionable workflows without requiring a “rip-and-replace” of legacy IT infrastructure?
Building a decoupled, non-invasive control plane
Enterprise ERPs are the ultimate rigid monoliths. Asking a customer to rebuild their foundation to adopt AI is a non-starter. We approached this by treating the intelligence layer as a decoupled, non-invasive control plane.
Event-driven and API-first architecture
Using an event-driven, API-first architecture, we listen to changes in the ERP without ever becoming a dependency for its core transactional logic. We use data contracts and schema normalization layers to bridge the gap between heterogeneous systems. This stateless design allows us to augment existing systems rather than replace them, injecting agility into environments that were previously locked-in.
Engineering for ‘Data Exhaust’ at Scale
How did you build a scalable ingestion pipeline to handle fragmented global supply chain data?
Designing for fragmented and messy data
Supply chain data is data exhaust it’s messy, fragmented, and arrives via legacy APIs and manual spreadsheets. Instead of hoping for clean data, we engineered for chaos.
From raw data to context-aware intelligence
We treat all incoming data as semi-structured signals. Our ingestion pipeline uses layered feature engineering to move from raw data to context-aware features. By prioritizing data lineage and traceability, we ensure every recommendation has a “paper trail” back to its source. This allows the system to scale across global networks without losing the granular context required for high-stakes decision-making.
The ‘Black Box’ Constraint: Building Explainable AI
In supply chain operations, a wrong recommendation can cost millions. How do you engineer “explainability” into the final output?
Making AI decisions interpretable
In our world, trust is a hard engineering constraint. If a human planner sees a recommendation but can’t see the logic, they won’t execute it. To solve the “black box” problem, we designed the system to expose interpretable intermediate states.
Combining machine learning with deterministic logic
We use a hybrid approach: we combine ML-driven probabilities with deterministic logic (hard rules and business constraints). This creates a transparent decision path. The system doesn’t just say “Reroute this shipment”; it shows the trade-offs, confidence scores, and specific constraints like port congestion or carrier costs. It’s a human-in-the-loop workflow where the AI justifies itself to the expert.
The Path Forward: Scaling Operational Velocity
As supply chains grow more complex, the bottleneck is no longer a lack of data it’s the human latency required to interpret it. Our mission is to bridge that gap. By moving the intelligence out of the monolith and into a scalable, explainable engine, we are giving planners the ability to move at the speed of their data, not the speed of their spreadsheets.
The goal isn’t just to predict the future of the supply chain; it’s to give enterprises the arc
Frequently Asked Questions
What is prescriptive AI in supply chain management?
Prescriptive AI in supply chain management goes beyond predicting outcomes and focuses on decision optimization. It evaluates multiple scenarios using constraints like cost, service levels, and capacity to recommend the best possible action.
How does prescriptive AI differ from predictive supply chain models?
Predictive models estimate what will happen, such as delays or demand changes. Prescriptive AI treats predictions as inputs and uses optimization and reasoning layers to determine the best action to take under real-world constraints.
How can AI work with existing ERP systems without replacing them?
AI systems can operate as a decoupled, API-first layer that listens to ERP events without interfering with core transactional systems. This allows businesses to enhance decision-making without replacing existing infrastructure.
How does AI handle fragmented supply chain data?
AI systems treat fragmented data as semi-structured signals and use ingestion pipelines with feature engineering, data lineage, and traceability to convert raw data into context-aware insights for decision-making.
How is explainable AI implemented in supply chain systems?
Explainable AI is implemented by combining machine learning probabilities with deterministic rules and constraints. This allows the system to show trade-offs, confidence levels, and decision logic behind each recommendation.
Why is decision optimization important in modern supply chains?
As supply chains become more complex, the bottleneck shifts from data availability to decision-making speed. Decision optimization reduces human latency and enables faster, more accurate actions based on real-time data.