Puran Water LLC

AI-native wastewater process engineering. Wastewater process engineering tools with schema'd state and deterministic simulation engines, designed for programmatic access by AI agents.

Puran Water builds the digital infrastructure for an industrial wastewater design-build-operate firm. The architecture rests on a central thesis: an AI-native industrial firm needs a properly schema'd data substrate first — not a memory system or RAG pipeline as the primary knowledge store.

That substrate comes from three sources: self-hosted enterprise OSS (each brings a Postgres-backed domain ontology), purpose-built engineering schemas (plant-state model, process unit types, ISA 5.1 instrumentation, DEXPI equipment classes, model credibility metadata), and custom domain schemas (procurement, process datasheets, bid specifications, compliance). All exposed via typed MCP tool surfaces.

External systems and human operators
            |
Communication / orchestration runtime
            |
Persona layer (role-scoped agents) + reusable skills
            |
MCP server layer (server codebases)       <-- this org
            |
Business systems + engineering engines

The MCP servers in this org are the tool layer — the typed interfaces between AI agents and deterministic engineering computation. They sit within a larger operating system that includes project management, CRM, procurement, compliance, and autonomous orchestration, documented in PuranOS-public.

The MCP ecosystem has grown to 10,000+ servers (as of early 2026). Many are developer and infrastructure oriented: databases, browsers, file systems, cloud APIs. While engineering-adjacent MCP servers exist for CAD tools, building energy simulation, and power grid modeling, the wastewater and chemical process engineering domain remains sparse.

This gap matters. Industrial process engineering has properties that generic AI tooling cannot address:

• Calculations must be deterministic and auditable. A pump sizing is not a language task. It is an engineering problem with a verifiable answer. LLM-generated numbers are not acceptable for engineering design — the calculation must be reproducible and traceable to published correlations or validated simulation models.

• Models chain across engines. A treatment train flows from biological treatment (QSDsan, mASM2d plant state) through reverse osmosis (WaterTAP, MCAS plant state). Sizing, simulating, and costing a flowsheet requires typed converters with provenance tracking at every handoff.

• Results carry credibility. A preliminary heuristic sizing, an uncalibrated simulation, and a calibrated dynamic simulation all produce a plant state. They have fundamentally different reliability. Every simulation result must carry explicit metadata: model status (calibrated/preliminary), decision grade (design/budgetary), and validation basis (bench-test/plant-data/literature/vendor/assumed).

• Formats must be machine-readable and standards-aligned. P&IDs as DEXPI XML, using DEXPI's vendor-neutral model and ISO 15926-aligned reference concepts. Process flows as SFILES text. Equipment tagged per ISA 5.1, hierarchy per ISA-95. Not PDFs. Not screenshots. Structured data that agents can read, write, diff, and validate.

These properties require purpose-built MCP servers — typed tool surfaces over physics-based simulation engines, not wrappers around chat APIs.

These tools implement the Model Context Protocol (MCP), providing structured JSON interfaces for deterministic engineering calculations. AI agents compose multi-step workflows by calling tools programmatically rather than requiring human operators to navigate graphical interfaces.

Engineering drawings follow a database-first architecture where machine-readable data models (DEXPI for P&IDs, SFILES for BFDs/PFDs) generate visualizations. This inverts the traditional CAD workflow, enabling version control via git and automated diff operations on the underlying data structures.

Engineering MCP servers generate Markdown reports with LaTeX equations for calculation traceability, Mermaid diagrams for process flowsheets, and Obsidian frontmatter for searchable metadata. This preserves Markdown as the version-controlled source of truth while enabling conversion to client-deliverable formats via Pandoc.

Most engineering MCP servers have been consolidated into the PuranOS monorepo. The public repos represent their standalone development at the point-in-time of consolidation and remain available as reference. Several were under active development and not yet production-ready — individual repo READMEs note their maturity status.

These engines maintain session state across agent interactions, support cross-engine handoffs via typed component-basis converters, and tag every result with model credibility metadata. A shared engineering library provides deterministic converters between component bases (mASM2d, MCAS, mADM1) with provenance tracking, along with shared Pydantic models for equipment items, credibility, and stream state. Inter-agent process data flows via filesystem JSON files conforming to the plant-state schema, validated by the StreamState model and an arithmetic mass balance checker — replacing the need for a dedicated state-management server.

freecad-pid-workbench represents the strategic direction for P&ID editing — fully open-source FreeCAD with DEXPI XML as the primary serialization, enabling git-based version control and AI-agent-accessible diagram editing.

• knowledge-base-mcp — Hybrid dense/sparse/rerank retrieval with Docling ingestion, Qdrant + FTS payloads, deterministic upsert tools.

Early explorations that validated MCP patterns with proprietary engineering tools:

• autocad-mcp — AutoCAD LT integration via AutoLISP. Informed the development of freecad-pid-workbench.
• mathcad-mcp — MathCAD Prime COM automation. Insights shaped the Markdown + LaTeX reporting approach.

The MCP servers are one layer. The full operating system — documented in PuranOS-public — includes:

• Schema'd state over memory — enterprise OSS schemas + engineering schemas + custom domain schemas as the primary knowledge substrate
• OpenProject as coordination substrate — shared board for human+AI task delegation, backed by agent coordination research
• Skills as captured expertise — reusable workflows that compound institutional knowledge
• First-class engineering computation — session-persistent engines with credibility metadata and typed cross-engine handoffs
• Standards alignment — DEXPI, ISA-95, ISA 5.1, CFIHOS, OPC UA
• Research backing — llmenron, StateFlow, Agent Workflow Memory, and counter-evidence

Start with Schema Over Memory to understand the central thesis, or Architecture Overview for the system structure.

• Puran Water LLC
• Hersh Kshetry, Founder and Principal Engineer
• Website: puranwater.com
• Contact: puranwater.com/contact
• Schedule a call: calendar.app.google/M1jzSdCB51sYiWux6