Fragmented signals
Operational data often sits across disconnected warehouse, labor, equipment, dock, and safety systems.
Business problem
When operations change in minutes, static dashboards are not enough
A demand surge, conveyor failure, labor shortage, dock constraint, or safety event can rapidly increase backlog and late-shipment risk. Managers must compare recovery actions while balancing throughput, overtime, equipment limits, cost, and worker safety.
High-speed trade-offs
The fastest response is not always the safest or most economical response.
AI without governance
An LLM alone should not calculate warehouse physics, approve overtime, or override operating constraints.
Solution
A decision-support layer between operational signals and human managers
FulfillTwin AI combines a simulated control tower, a digital-twin scenario laboratory, traditional machine learning, local policy retrieval, specialist agents, deterministic plan scoring, and human approval.
Simulated operational monitoring
Generates representative event streams for portfolio demonstration and provides a control-tower pattern ready for production data connectors.
Disruption mapping
Maps events such as demand surges, labor shortages, and equipment failures into controlled digital-twin scenarios.
ML impact estimates
Estimates future backlog and SLA-breach risk, identifies the operating regime, and flags unusual states.
Seven-agent council
Specialist agents independently assess demand, labor, equipment, docks, energy, safety, and financial impact.
Deterministic optimization
Scores candidate plans using explicit cost, backlog reduction, throughput, overtime, and safety constraints.
Auditable decisions
Writes scenarios, forecasts, recommendations, and approvals to an atomic, thread-safe JSON decision store.
Project evidence
What is implemented, what is demonstrated, and what still requires production data
This section separates working technical components from future production validation so employers can evaluate the project accurately.
Data
Current: synthetic fulfillment scenarios generated at startup.
- Order volume and backlog
- Absenteeism and labor capacity
- Equipment, dock, safety, and energy signals
Models
Implemented: a mixed traditional-ML pipeline.
- XGBoost regressor
- XGBoost classifier
- K-means clustering
- Isolation Forest
Evaluation
Available in Model Ops: demonstration metrics, model information, retraining controls, and decision-memory inspection.
Synthetic-model results are engineering evidence, not claims of live warehouse accuracy.
Production limitation
Required next: replace the synthetic generator with site-specific historical events, define operational baselines, retrain, validate by facility, and move audit records to persistent production storage.
Technical differentiation
Not just another LLM wrapper
The LLM is an optional communication layer. The operational recommendation is grounded in traditional ML outputs, deterministic rules, local retrieval, and transparent plan scoring.
| Capability | Implementation | Role in the decision |
|---|---|---|
| Future backlog estimate | XGBoost regressor | Produces the numerical backlog forecast used by agents and plan scoring. |
| SLA-breach risk | XGBoost classifier | Estimates late-shipment risk for the simulated operating state. |
| Operating regime | K-means clustering | Groups the scenario into an interpretable operating pattern. |
| Out-of-distribution warning | Isolation Forest | Flags scenarios that differ substantially from the training distribution. |
| Safety and operating constraints | Deterministic expert system | Blocks or flags actions that violate encoded guardrails. |
| Policy evidence | Local TF-IDF RAG over playbooks | Retrieves relevant guidance without making the LLM the source of truth. |
| Recovery-plan selection | Transparent deterministic arbiter | Compares candidate interventions using explicit costs and operational effects. |
| Executive explanation | Local, Groq, or Gemini provider | Turns the already-selected plan into a manager-readable brief. |
Worked demonstration
Example: demand surge and constrained recovery
A representative Scenario Lab run shows how the system moves from model outputs to a governed recovery recommendation.
Service-first recovery selected for a high-risk simulated event
The digital twin estimated a severe backlog and full SLA-breach risk. The agent council generated alternatives, and the deterministic arbiter selected the lowest-scored viable plan under the configured rules.
Selected actions in the displayed run included moving 15 associates, authorizing 4 overtime hours, applying a 22% release throttle, and requiring human approval before execution. These values illustrate system behavior; they are not production performance guarantees.
Decision flow
From warehouse signal to manager-approved action
Step 1
Operational inputs
Simulated or future site-integrated warehouse signals.
Step 2
ML and digital twin
Backlog, SLA risk, anomaly, and operating-regime outputs.
Step 3
Rules, RAG, and agents
Guardrails, policy evidence, and specialist assessments.
Step 4
Deterministic arbiter
Transparent comparison of cost, safety, and recovery impact.
Step 5
Human approval and audit
Manager control, executive brief, and recorded decision history.
End-to-end architecture linking operational inputs, traditional ML, expert guardrails, local retrieval, the seven-agent council, deterministic optimization, human approval, and the control-tower feedback loop.
Engineering stack
Built as a deployable Streamlit and Flask application
Application layer
Streamlit multipage UI, Flask JSON API, health endpoint, Docker startup, and Railway deployment configuration.
Intelligence layer
XGBoost, scikit-learn, local expert rules, TF-IDF retrieval, seven specialist agents, and optional Groq or Gemini narratives.
Governance layer
Human-in-the-loop approval, atomic JSON decision memory, local fallback behavior, model operations, and test coverage.
Python
Streamlit
Flask
XGBoost
scikit-learn
TF-IDF RAG
Groq
Gemini
JSON Audit Memory
Docker
Railway
pytest
Responsible deployment
Designed as advisory decision support—not autonomous workforce control
FulfillTwin AI does not authorize material interventions on its own. Overtime, major labor reassignment, and high-risk recovery actions remain subject to explicit human review.
Deterministic guardrails
Encoded rules check capacity, safety, and operating constraints before a narrative is produced.
Provider-independent fallback
If an external LLM provider is unavailable, the application can return a deterministic local brief.
Traceable recommendations
The application records the scenario, model outputs, candidate plans, final recommendation, and approval state.
Production hardening would include authenticated roles, persistent database or object storage, facility-specific validation, monitoring and drift detection, secrets management, incident-retention policies, and integration testing against real source systems.