How Nexaminds helped EverestLabs build and deploy a governed conversational intelligence layer that makes complex industrial data easier to access and act on.
Industrial operations generate enormous volumes of data.
Production systems, robotics, sensors, sorting equipment, quality-control systems, and operational databases continuously capture information about throughput, runtime, material flow, downtime, performance, and efficiency.
But having the data does not necessarily make it easy to use.
For an operations team, answering a seemingly simple question can still require finding the right dashboard, identifying the correct facility or production line, selecting the right time window, understanding internal naming conventions, or asking an analyst for help.
What if an operator could simply ask:
What changed yesterday?
And then follow with:
Which production line contributed most to that change?
That is the experience Nexaminds and EverestLabs set out to enable using Claude and the Model Context Protocol (MCP).
The resulting architecture now lives within EverestLabs’ operational production environment, providing a governed conversational layer over industrial intelligence while keeping operational definitions, permissions, calculations, and data access within the systems designed to control them.
Partner and Customer Background
Claude Partner: Nexaminds
About Nexaminds
Nexaminds specializes in providing cutting-edge AI consulting services to help transform businesses and navigate the complexities of diverse legacy and AI technologies to increase efficiency, reduce costs, and unlock new revenue streams. Nexaminds helps its customers in the following areas:
Machine Learning and Artificial Intelligence: Transform and automate your business with intelligent AI and machine learning solutions, gain valuable data-driven insights, improve decision-making, enhance customer experiences driving efficiency, growth, and competitive advantage.
Application Modernization: Upgrade to modern architectures and technologies, ensuring future-proof applications.
Cloud Transformation: Deploy, manage, and scale your infrastructure with ease.
Agile Development: Software Development services powered by Cloud Native Engineering, CI/CD, DevOps, Test Automation, and iterative delivery leveraging Scrum, Kaban, and SAFe to accelerate secure, high-quality, software releases.
Data Engineering: Construct scalable data solutions. We optimize data ingestion, processing, and storage for peak performance.
Partner Customer: EverestLabs.AI
About the Partner Customer
Founded in 2018, EverestLabs is an award-winning Physical AI company building intelligent solutions for the industrial economy. Our first market is the materials economy, where we apply agentic AI and robotics solutions to transform how materials are recovered, processed, and returned to productive use.
The materials economy — spanning metals, plastics, and fiber from both recycled (~$1T) and virgin (~$2.5T) source - forms a critical foundation of the $6T manufacturing economy across the U.S. and EU.
Traction: 250+ systems (AI cameras and robots) deployed in recycling plants (also called material recovery facilities) and reclaimers (companies that buy recycled bales of aluminum cans and convert them into recycled aluminium sheets)
100B data points about packaging, nearly a trillion inferences, edge inference in < 15 milliseconds with > 95% accuracy.
Core IP & Moat: Physical AI Software Stack from the ground up that encompasses Autonomy, Intelligence, and Hardware Layer.
EverestLabs team has developed Purpose Built Large Vision Models (LVM) with data coming from hundreds of cameras & adapted Off the Shelf Robots to the Materials Industry with the following moat:
Proprietary data advantage (The training data doesn’t exist on the public internet nor VLAs will work for industrial settings with highest variability)
Full-Stack Platform For Customers (See, Learn, Act, Orchestrate -delivered as one system),
Realtime Agentic Stack (Realtime data analysis & orchestration engine)
Problem Statement
Industries in the materials economy space produce enormous amounts of data. This data comes from various equipment such as optical sorters, conveyors, SCADA systems, PLCs, IoT and various systems. EverestLabs’ team is processing millions of data points from various systems across its fleet. One of their goals is to make this data readily available to the plant operators as an agentic conversational agent to make informed decisions proactively. After extensively speaking with dozens of customers and plant operators, EverestLabs team has developed the product and solution road map from the ground up. This is where they collaborated with us to implement their agentic vision for industrial plants using Claude + EverestLabs AI + Nexaminds’ supporting programmers to speed up the launch of the platform.
Why Now and Advantages
For the first time, a seamless but disparate technologies and solutions exist to deliver results that were not possible previously.
By combining a new chat agent with the core IP of EverestLabs team across the whole plant, plant operators now truly see the power of agentic AI to increase throughput & utilization, improve recovery and end product quality, reduce downtime and track KPIs in real time.
Data Privacy, Safety, and Security
No customer data was shared with Nexaminds or Anthropic.
Claude did not train on EverestLabs customer data or plant data.
All customer data, LVM AI models, machine data resides within EverestLabs ecosystem.
All IP rights and work-product reside with EverestLabs.
The Challenge: Making Industrial Intelligence Easier to Access
EverestLabs already brings together AI, robotics, equipment intelligence, and operational data to help recycling facilities understand and optimize what is happening across their operations.
The opportunity was not to create another source of industrial data.
It was to make the intelligence already available across those systems easier for people to interact with.
That sounds straightforward until we consider a question such as:
What was the performance of the output line yesterday?
To the user, this is a simple operational question.
To the underlying system, it contains several important details that must be resolved correctly:
Which facility does the user mean?
What does “output line” represent at that facility?
Which equipment belongs to that process?
Which timezone defines “yesterday”?
Is the user authorized to access that facility?
Which governed analytical source defines the requested performance metric?
A language model can understand what the user is trying to accomplish.
But it should not independently invent operational definitions, equipment mappings, permissions, or analytical formulas.
That distinction became the foundation of the solution.
Claude handles language, reasoning, conversational context, tool selection, and response synthesis.
The MCP and operational layers handle controlled execution, data access, mappings, permissions, and deterministic business logic.
What We Built: A Governed Conversational Layer for Industrial Operations
Through our collaboration with EverestLabs, we built a Claude-integrated MCP architecture that provides a natural-language interaction layer over industrial operational intelligence.
Conceptually, the system follows this flow:
Industrial Operations → EverestLabs Operational Intelligence → Governed MCP Capabilities → Claude → Operations Teams
Each layer has a clear responsibility.
EverestLabs’ Physical AI, robotics, operational systems, and data platform provide the underlying industrial intelligence.
The MCP layer exposes selected analytical capabilities through controlled tools.
Claude interprets the user’s question, determines which capability is required, invokes the appropriate tools, and turns the resulting structured information into an understandable response.
Nexaminds designed and implemented the conversational and orchestration architecture connecting those components.
This allows an operations user to ask questions such as:
How many hours was the line active yesterday?
What was the throughput during the morning shift?
Compare material composition across the last seven complete days.
And then continue naturally:
Now compare only the morning shifts.
Which day had the largest decline?
Instead of requiring the user to translate each question into database logic, Claude maintains the conversational context while the governed backend remains responsible for producing the analytical result.
Why Claude + MCP
Claude and MCP solve different parts of the problem.
Claude provides the reasoning and conversational layer.
Industrial users speak in operational language - not database schemas.
Terms such as morning shift, active time, throughput, material composition, or last complete week need to be interpreted in the context of the user’s request.
Claude can understand that intent, determine which analytical capability is appropriate, maintain context across follow-up questions, and synthesize structured results into a useful response.
MCP provides the connection to governed capabilities.
Rather than giving Claude unrestricted access to an operational database, the MCP server exposes specifically designed tools for trusted analytical functions.
Claude determines what needs to be done.
The MCP and operational layers determine how it is executed.
This separation allows the conversational experience to benefit from Claude’s reasoning while keeping operational logic in deterministic systems.
The High-Level Architecture
At a high level:
Industrial Data Sources → Operational Data Platform → MCP Analytics Layer → Claude → User Interface
Industrial Data Sources: Sensor, equipment, robotics, MES/ERP, and quality-system data
Governed Data Platform: Cloud-hosted relational data warehouse and operational data services
MCP Analytics Server: Governed orchestration and tool-calling layer exposing operational analytics capabilities
Claude LLM Sonnet: Reasoning, tool selection, conversational context, and response synthesis
Operational Interfaces: Grafana dashboards and conversational analytics interfaces
The request flow works as follows:
Industrial systems generate plant, equipment, production, and material data.
EverestLabs’ operational platform transforms those signals into trusted operational intelligence.
The MCP layer exposes selected analytical capabilities as governed tools.
Claude interprets a natural-language question and selects the appropriate capability.
The backend resolves the required facility, equipment, time, and access context.
The analytical operation is executed against the governed source.
Structured results are returned to Claude.
Claude presents the answer and maintains context for the next question.
The most important architectural principle is simple:
Claude reasons over industrial capabilities; it does not replace the systems responsible for defining them.
That distinction helped us move the experience from a conversational prototype toward a production capability.
Four Engineering Decisions That Made It Production-Ready
Connecting Claude to an MCP server was only the beginning.
For an industrial environment, reliability depended on how the capabilities behind that interface were designed.
Four engineering decisions became particularly important.
1. Domain-Specific Tools Instead of Generic SQL
One of our key decisions was to avoid exposing a generic “run SQL” capability to Claude.
While flexible, that approach would require the model to correctly understand database schemas, joins, naming conventions, units, metric definitions, and access boundaries.
Instead, the MCP server exposes domain-specific analytical capabilities.
For example, a throughput tool may receive:
Facility
Process or production line
Start time
End time
The backend then performs the required logic:
Resolve the facility
Validate access
Map the process to the correct equipment
Resolve the operational timezone
Select the governed analytical source
Execute the approved calculation
Apply required transformations
Return structured results
Claude does not need to know the physical database schema.
It needs to understand when the capability should be used and what information is needed to invoke it correctly.
2. Operational Context Lives Outside the Model
Industrial terminology changes across facilities.
A process name at one plant may map to one equipment identifier, while the same business term at another plant may map to several pieces of equipment.
We therefore keep facility-specific context in a governed operational layer rather than embedding every mapping inside a growing prompt.
That context can include:
Facility-to-timezone mappings
Facility-to-production-line mappings
Process-to-equipment mappings
Metric definitions and units
User-access permissions
This creates a clean separation:
Claude instructions define behavior.
MCP defines available capabilities.
The context layer defines facility-specific operational metadata.
EverestLabs’ operational platform remains the source of analytical truth.
This structure also makes it easier to expand the system as additional facilities and capabilities are introduced.
3. Time Is Treated as Operational Context
A request as simple as:
What happened yesterday?
cannot be answered reliably without knowing the facility’s timezone.
Operational databases may store timestamps differently from the local time used by plant operators.
The system therefore resolves the facility and its timezone before interpreting expressions such as:
Today
Yesterday
Morning shift
Last seven days
Last seven complete days
Those time windows are handled explicitly by the operational layer rather than leaving their definition entirely to the language model.
It is a small architectural detail with a significant effect on analytical reliability.
4. Authorization Is Enforced Below Claude
A conversational interface should not bypass the governance of the systems behind it.
Authorization therefore remains server-controlled.
The backend determines which operational context a verified user is permitted to access before analytical data is returned.
If a user is not authorized to access a facility or capability, the request is rejected before the data is retrieved.
The governing principle is:
Claude can behave in an authorization-aware way, but Claude is not the authorization system.
The same philosophy applies to structured outputs.
MCP capabilities return explicit operational results — including relevant facility, process, time window, unit, and metric information — which Claude can then explain conversationally without changing the underlying analytical truth.
What Happens When an Operator Asks a Question
Consider:
Compare active hours for the packaging lines over the last seven complete days.
Claude identifies the user’s intent:
Metric: active hours
Process: packaging lines
Time period: seven complete local days
Operation: comparison
It then selects the appropriate MCP capability.
Behind the interface, the operational layer:
Resolves the user’s permitted facility.
Maps “packaging lines” to the relevant equipment.
Retrieves the facility timezone.
Resolves the requested seven complete local days.
Queries the governed analytical source.
Returns structured results by line and date.
Claude receives those results and turns them into a concise operational response.
The user can then continue:
Which day performed the worst?
or:
Now show only the morning shift.
From the user’s perspective, this is one analytical conversation.
Underneath it, each question is still resolved through controlled operational capabilities.
From Prototype to Production
The Claude-integrated MCP architecture is now live within EverestLabs’ operational production environment.
Moving into production required preserving the operational intelligence and governance already present within the EverestLabs ecosystem while providing a reusable way for Claude to interact with it.
The resulting system keeps:
reasoning and conversation with Claude,
while keeping:
operational definitions, mappings, permissions, calculations, and data access within governed backend systems.
That separation is important for scalability.
The conversational experience can evolve without moving plant-specific business logic into the model.
At the same time, EverestLabs can evolve its facilities, equipment mappings, metrics, and analytical capabilities behind a stable MCP interface.
This creates a foundation that can expand beyond a single question, tool, or facility.
Why the Collaboration Model Matters
The implementation combines complementary capabilities across EverestLabs, Nexaminds, and Claude.
EverestLabs provides the industrial intelligence.
Its Physical AI platform, robotics, operational systems, domain expertise, and plant intelligence form the foundation of the experience.
Nexaminds provides the enterprise AI integration and orchestration layer.
Our work connects Claude with governed operational capabilities through MCP and implements the context, tool orchestration, and production architecture required to make those interactions reliable.
Claude provides the reasoning and conversational layer.
Claude interprets operational questions, selects the appropriate governed capabilities, maintains conversational context, and turns structured analytical results into answers users can act on.
Keeping these responsibilities distinct is what allows the system to combine powerful conversational reasoning with the reliability required in an industrial environment.
What Comes Next
With a governed MCP foundation in place, conversational analytics can evolve into increasingly agentic operational workflows.
For example, the same architecture can support:
Anomaly investigation: identifying unusual changes in throughput, runtime, material composition, or equipment performance and helping users investigate them conversationally.
Root-cause analysis: combining multiple governed capabilities to evaluate whether a change in performance coincided with downtime, material changes, or other operational conditions.
Operational summaries: generating concise shift or management summaries based on consistent, governed operational metrics.
The important point is that these capabilities can build on the same architecture already used for conversational analytics rather than requiring an entirely new system.
Final Thoughts
Measurable Operational Impact
The Claude- and MCP-powered analytics layer transforms how operators interact with complex plant data, replacing multi-step dashboard and manual analysis workflows with a natural-language conversational experience.
60% reduction in analysis time: High-level operational analysis that previously required navigating dashboards, applying filters, correlating multiple metrics and interpreting results can now be performed through a natural-language chat interface. By using Claude to understand the operator’s question and MCP to securely retrieve the relevant plant data, the solution reduces time-to-insight by approximately 60% for common analytical workflows.
Faster insights supporting 20-30% greater facility capacity: The conversational analytics layer gives operators faster access to throughput, belt utilization, material-flow and performance data, allowing them to identify bottlenecks, investigate underutilized capacity and make optimization decisions more quickly. These capabilities support Navigator’s broader load-balancing and process-optimization workflows, which EverestLabs reports can increase facility capacity by 20-30%.
Accelerated analysis supporting $1M in downtime-prevention value: Through Claude and MCP, operators can quickly investigate equipment performance, uptime, anomalies and historical operating conditions using a single conversational interface. Faster access to this context helps teams understand and act on predictive-maintenance signals, supporting Navigator’s broader downtime-prevention capabilities, which EverestLabs estimates can represent approximately $1M in annual value.
Together, Claude and MCP reduce the analytical effort required to turn complex plant data into actionable intelligence, shortening the path from question → analysis → insight → operational action.
Industrial operations are becoming increasingly intelligent.
AI vision systems can understand what is moving across production lines.
Robotics can act on that intelligence.
Sensors and equipment continuously generate information about what is happening inside a facility.
Operational platforms can turn those signals into intelligence about production, materials, quality, uptime, and performance.
The next opportunity is making that intelligence easier for people to use.
The collaboration between EverestLabs and Nexaminds demonstrates one approach: combining EverestLabs’ industrial intelligence with Claude and MCP to create a governed conversational interface for operations.
Instead of requiring an operator to know where a metric lives or how the underlying analytical system works, the interaction can begin with a simple question:
What changed yesterday?
And continue naturally:
Which part of the process contributed most?
Behind that conversation sits a deliberate architecture:
EverestLabs’ systems generate and govern operational intelligence.
MCP exposes controlled analytical capabilities.
Deterministic services enforce operational context and rules.
Claude reasons across the results and maintains the conversation.
People remain responsible for the decisions.
That is where conversational AI becomes particularly powerful in industrial environments:
not as a replacement for operational intelligence, but as the interface that makes that intelligence dramatically easier to use.
Summary
Before the Solution
EverestLabs had access to rich operational data across its recycling facilities, but extracting useful insights often required users to navigate multiple dashboards, understand plant-specific data structures, apply filters and manually correlate metrics. This created friction between having the data available and being able to quickly turn it into actionable operational insight.
The Customer Problem
EverestLabs wanted to make complex plant analytics easier and faster for operators and business users without requiring them to understand the underlying data models, SQL queries or dashboard structures. The goal was to enable users to ask operational questions naturally and receive relevant, governed answers grounded in real plant data.
The Claude + MCP Solution
Nexaminds worked with EverestLabs to build a Claude-powered conversational analytics layer using Model Context Protocol (MCP). Claude serves as the reasoning and conversational interface, while MCP securely connects it to governed operational tools and plant data.
When a user asks a question in natural language, Claude interprets the request, identifies the appropriate analytical tools, resolves the relevant plant, equipment and time context, retrieves the required operational data through MCP, and presents the result in an understandable, actionable format.
This allows users to move from complex dashboard-driven analysis to a conversational workflow:
Question → Claude → MCP Tools → Governed Plant Data → Insight → Action
Measurable Impact
The natural-language analytics experience reduces the time required for common high-level operational analysis by 60%, allowing users to reach insights faster than through traditional multi-step dashboard and manual analysis workflows.
By accelerating access to throughput, utilization, material-flow, equipment and performance data, the Claude and MCP layer also supports the broader Navigator optimization capabilities that EverestLabs reports can contribute to 20-30% increases in facility capacity and $1M in annual value from downtime prevention.
Overall, the solution shortens the path from operational data to decision-making by giving plant teams a faster, simpler and more intuitive way to interact with their data.