Engineering Computation
Deterministic Tools. Typed Interfaces.
MCP servers wrap peer-reviewed open-source engines and expose them as typed, auditable tools. Every input is validated. Every output is structured. Every call is logged.
How Engineering Computation Works in PuranOS
In most engineering firms, simulation models live in notebooks, Excel spreadsheets, or desktop applications disconnected from project state. An engineer runs a model, copies results into a report, and the model becomes stale the moment project parameters change.
PuranOS treats engineering computation differently. Each engine is API-first (not GUI-first), session-persistent across agent interactions, and produces results with explicit credibility metadata.
Agents don't interact with these tools through a chat interface. They are triggered by OpenProject task assignments, scheduled automation, or inbound emails — and they call engineering tools as part of structured skill execution.
Core Engineering Engines
QSDsan
University of Illinois · Anaerobic digestion and dynamic LCA
IWA ADM1 anaerobic digestion, dynamic mass/energy balance, life-cycle assessment. Session-persistent simulations with model credibility metadata on every result. Steady-state activated sludge biology is now solved in WaterTAP via ASM2D — QSDsan handles dynamic regimes, anaerobic systems, and LCA studies where time-resolved component tracking is required.
WaterTAP
NAWI / DOE · Steady-state biology, membranes, separation, costing
Canonical steady-state engine: ASM2D biological treatment (MLE, MBR-MLE, MBR-MLE_POST) with SRT-as-constraint formulation reaches IPOPT optimal in seconds; MLSS converges to design target. Also covers RO, NF, ED, crystallizer, evaporator, and exact-Hessian costed optimization. Cross-engine handoffs from QSDsan and 31-component stream-state via typed converters with provenance tracking.
PHREEQC
USGS · Chemical equilibrium
Precipitation, dosing optimization, scaling potential analysis. Multi-database support (phreeqc.dat, minteq.dat, llnl.dat, pitzer.dat).
CoolProp + fluids
Open source · Fluid mechanics
120+ fluid properties, pipe sizing, pump TDH, control valve sizing (IEC 60534), blower power calculations.
Heat Transfer
Caleb Bell et al. · Thermal engineering
Heat exchangers, insulation design, heat trace. 390+ material conductivities (VDI/ASHRAE). Built on ht, chemicals, fluids, thermo.
Corrosion Engineering
USGS + Caleb Bell · Corrosion analysis
Sweet/sour corrosion, galvanic analysis, pitting risk assessment. Built on phreeqpython, fluids, ht.
Engineering (Diagrams & Tagging)
Process Intelligence Research · P&IDs and BFDs
DEXPI P&IDs, SFILES BFDs/PFDs, ISA 5.1 instrument tagging. Built on pyDEXPI, SFILES2, networkx.
Equipment Identity Registry
ISO 14224 · Cross-system equipment identity
Canonical equipment identity: functional positions (ISA 5.1 tag + UUID) linked to physical asset instances (CMMS ID + serial). Lifecycle tracking from design through decommissioning. The governed master identity that connects P&ID, CMMS, procurement, inventory, and project management.
Knowledge Wiki (wiki-graph)
File-backed markdown vault · Wikilink graph traversal
Agent-maintained wiki for context that does not fit in schema'd databases: meeting synthesis, email threads, research paper summaries, design rationale, competitive intelligence, and commissioning lessons. Drop-flow agents write raw captures to dated channel folders; a knowledgebase agent reconciles raw into linked compiled articles nightly. Served by wiki-graph (fork of mcpvault): 15 file I/O tools + 8 graph tools (search_traverse, backlinks, fragment, find_path, and more) + health and safe-rename primitives + MCP resources for ambient context. Agent-only access — no human UI. Complements the knowledge-base server (structured vector search) by providing narrative context alongside structured retrieval.
RLM Bridge + ArtifactStore
Codex-backed reasoning · Multi-thousand-document corpus synthesis
Cross-corpus reasoning over tender documents at multi-thousand-page scale. Documents extracted via a Modal-hosted Docling endpoint, indexed by content hash with a docling cache guardrail, and reasoned over by the RLM (Long-context Language Model) bridge with Codex as the backend tool-use loop. A companion ArtifactStore MCP lets a single reasoning pass write structured artifacts across multiple databases in one transaction, replacing brittle multi-prompt extraction chains.
Ontology-MCP (universal read layer)
Cross-database read · Typed traversal across the link catalog
Single MCP surface for read-only access across all 7 Postgres databases. Domain servers (engineering, procurement, compliance, equipment-identity, contractor-management, project-controls, project-finance) remain write-only on their tables; ontology-MCP exposes ~50 read tools that traverse the typed link catalog, resolve cross-DB references, and assemble unified entity views without round-trips. Eliminates the per-server tool duplication that grew as the domain catalog expanded.
OpenProject + Hatchet
Canonical task ledger · Durable workflow execution
OpenProject (self-hosted, Enterprise features unlocked) is the canonical task ledger: 18 work-package types and 61 custom fields, with agent-state, equipment-tag, and interop fields woven into the type system. Hatchet provides durable workflow execution with explicit retry, timeout, and idempotency semantics — the tender response workflow, transcript reconciler, and communication-agent all run end-to-end on Hatchet under OpenProject's task model. An openproject-mcp server exposes typed CRUD; a hatchet-mcp server lets agents start, observe, and cancel workflow runs from inside a skill.
All servers run on fastmcp ≥ 3.0 with the full primitive set — tools, resources, elicitation, and prompts. Persistent sessions across calls, structured cancellation, and MCP-native error envelopes.
Cross-Engine Handoffs
A QSDsan effluent composition becomes a WaterTAP influent specification through typed converters — with unit conversion, component mapping, and provenance from the source model.
Multiple component bases are supported (mASM2d for activated sludge, MCAS for WaterTAP solutes, mADM1 for anaerobic models) with typed converters between them. The handoff is not a copy-paste — it carries the full lineage of how the data was produced.
Converters preserve balance reports, track estimated fields explicitly, and maintain source breakdown so credibility does not silently degrade across engine boundaries. When a stream crosses from one engine to another, the downstream consumer sees exactly which fields are measured, which are estimated, and what assumptions the conversion applied.
Model Credibility Metadata
An agent-generated sizing is not automatically "engineering grade." Every simulation result carries explicit credibility metadata:
Model status
validated · calibrated · heuristic · preliminary · stub
Decision grade
design · budgetary · screening · order-of-magnitude
Validation basis
bench-tested · plant-data · literature · vendor · assumed
As models improve, credibility metadata becomes more important, not less. A more capable model produces more convincing but still physically wrong answers when applied outside its valid range. An LLM 100x more capable still cannot reliably compute steady-state effluent composition — it lacks numerical precision, validated kinetic parameters, and iterative solver architecture. The credibility metadata is what distinguishes a deterministic engine result from an LLM-generated number.
The credibility metadata is what allows a PE reviewer to know what level of trust to place in a result.
In Practice: RO System Sizing
A concrete example of how engines, typed handoffs, and credibility metadata work together:
Query stream-state schema
Agent queries the stream-state schema for current stream composition — 31-component mASM2d basis with flow, temperature, pressure, pH, and component concentrations.
Solve steady-state activated sludge
WaterTAP solves the ASM2D biological treatment flowsheet (MLE, MBR-MLE, or MBR-MLE_POST) with SRT-as-constraint formulation. Prior to the asset becoming operational, simulations are always uncalibrated — based on literature or bench-scale testing. Results carry credibility metadata: model_status: preliminary, validation_basis: literature.
Typed converter maps effluent to RO influent
Converter transforms ASM2D effluent (mASM2d basis) to WaterTAP MCAS influent for the downstream RO stage with provenance tracking and estimated-field annotation. The handoff carries the full balance report so credibility does not silently degrade.
Run RO simulation
WaterTAP runs the reverse osmosis simulation — permeate quality, energy demand, capital cost estimate. Results carry their own credibility metadata.
PE reviewer sees full provenance chain
Final results carry credibility from both engines. The reviewer sees which numbers are from calibrated models versus screening-level estimates, and which fields were measured versus estimated at each handoff boundary.
Every step operates on schema'd data. There is no retrieval ambiguity, no garbage context, and no question about what the data means.
These engines are open-source. The MCP server wrappers and architecture are documented on GitHub.