Liminal Decision AI

Inspiration

Global supply chains lose over $1.6 Trillion annually due to sudden port blockades, factory lockdowns, and SLA breach penalties. When disruptions strike, traditional predictive AI fails because it only detects surface correlations rather than true root causes. Standard LLMs often hallucinate unviable recommendations or violate accounting laws. We built Liminal Decision AI to create a self-driving Causal AI platform that empowers enterprise managers to move from reactive panics to provably safe, risk-bounded action.

What it does

Liminal Decision AI is an autonomous Causal AI platform acting as a self-driving GPS for enterprise supply chains:

• Instant Data Retrieval: Embedded DuckDB & Apache Arrow memory engines query 50,000 multi-echelon disruption records in $< 2\text{ms}$.
• Root Cause & Treatment Effect Estimation: Uses Judea Pearl Structural Causal Models (SCMs) and EconML Double Machine Learning (DML) to compute unbiased Average Treatment Effects (ATE) for rerouting strategies.
• 95% Conformal Risk Control (CRC): Provides distribution-free 95% mathematical risk bounds ($\alpha = 0.05$) guaranteeing recommendations avoid financial loss and SLA penalties.
• Neuro-Symbolic Guardrails: Integrates Google Gemini 3.5 Flash with dynamic database-backed rules enforcing profit margin floors ($\ge 12\%$) and ESG carbon caps ($\le 4.5\text{ MT}$).

How we built it

• Layer 1 (Data Engine): PostgreSQL EF Core DB + Embedded DuckDB / Parquet columnar memory storage.
• Layer 2 (Causal ML Engine): Python & C# Causal Engine implementing Judea Pearl Abduction, EconML LinearDML, and 95% Conformal Risk Control.
• Layer 3 (Agentic AI & Symbolic Guardrails): Google Gemini 3.5 Flash Agent coupled with dynamic database-backed rule verification.
• Layer 4 (Frontend Studio): React 18 + Vite modern UI with Cytoscape/Mermaid network topology visualizer and CIO Candidate Exam Studio.

Challenges we ran into

• Orthogonal Residualization: Implementing Double Machine Learning (DML) required residualizing confounders out of both outcome $Y$ and treatment $T$ simultaneously.
• Zero-Copy Memory Ingestion: Connecting Python ML models to C# .NET 8 Web API via DuckDB in-process Arrow memory sharing without disk I/O bottlenecks.
• Dynamic Guardrail Scoping: Ensuring database-backed rule thresholds can be modified at runtime without recompiling backend C# logic.

Accomplishments that we're proud of

• Achieving sub-2ms query performance across 50,000 multi-echelon disruption records.
• Formulating a mathematically provable 95% Conformal Risk Control bound for supply chain interventions.
• Building a full-stack enterprise platform with 0 build warnings and 0 compilation errors.

What we learned

• How Structural Causal Models (SCMs) solve the fundamental flaw of correlation-based AI models in high-risk environments.
• Practical implementation of Neuro-Symbolic AI, combining flexible neural reasoning with deterministic business constraints.

What's next for Liminal Decision AI

• Multi-region streaming ingestion via Apache Kafka / Redpanda.
• Expanding Graph Neural Networks (GNNs / PyTorch Geometric) for deep multi-echelon tier-3 supplier risk mapping.

Built With

• apache-arrow
• c#
• causal-inference
• docker
• dotnet-core
• dowhy
• duckdb
• econml
• google-gemini-api
• machine-learning
• postgresql
• python
• react
• rest-api
• tailwindcss
• terraform
• vite