PWM-1F: Project World Foundational Model

Temporal, metric-first AI for agents that work in the project domain.

What PWM-1F does

A foundation model that treats projects as evolving systems

PWM-1F is a domain foundation model that acts as a project specialist inside your system. It is trained to read multi-week histories of project metrics, decide what information it needs, and produce project-aware analysis that your products and agents can use directly.

For any project related request, PWM-1F can be given:

  • The user’s question or task
  • The relevant project or portfolio scope
  • Access to project tools such as metrics APIs, risk registers and schedule views

Within that boundary, PWM-1F interprets the query in project terms, plans the steps it needs to take, calls the tools you expose, consumes the returned data, and then provides:

  • A clear view of project or portfolio health and trajectory
  • Positive and negative drivers behind schedule and margin movement
  • Short horizon risk and stability signals
  • Recommended focus areas or next steps
ai for project management diagram metrics
input

Inputs: metric histories and targeted context

PWM-1F is designed to work with simple tabular inputs plus optional context:

  • Metric history: Multi-week histories for each project, provided in a compact TSV. Metrics cover schedule signals, cost and margin indicators, throughput and workload, and risk or stability markers.
  • Project context (optional): Data such as project type, phase, portfolio tags, customer segment or region.
  • Targeted artefacts (optional): When the surrounding agent decides more detail is needed, it can pass selected risks, schedule slices or work package summaries as additional context.

PlanVector provides a metric dictionary and input specification so that platforms, enterprises and integrators can map their existing data into this format.

output

Outputs: agent-ready project analysis

PWM-1F returns natural language analyses that are ready for consumption by end users or downstream systems. Typical responses include:

  • Overall project status framed in clear language
  • Explanation of key positive and negative drivers, with references to time windows and signals
  • Short horizon view on delay and margin risk
  • Recommended areas of focus for the next period

When requested, responses can also contain a structured summary section that agents or analytics layers can parse.

What makes PWM-1F different

Metric-first and temporal

PWM-1F is trained to read sequences of metrics over time, rather than single snapshots or only documents. It learns patterns in how signals move together across weeks and what those patterns tend to mean.

Sensitivity to project dynamics

Through training on synthetic project histories, the model develops temporal causal sensitivity. It learns which changes commonly precede shifts in health, delay risk or margin stability.

Synthetic, domain-grounded training

The training data consists of synthetic project histories generated to behave like real portfolios. Real datasets are used at an aggregate and for validation. This keeps customer data out of the base model while allowing rich domain behavior.

Cross-platform by design

Because PWM-1F reasons over domain metrics rather than one platform’s schema, other systems can map their metrics into the PlanVector metric set and obtain consistent behavior.

How PWM-1F fits into agent architectures

The project brain inside your agent stack

PWM-1F is designed to work in more than one agent pattern. At a high level, there are two main ways to use it, plus hybrids.

1. Single-model agent: PWM-1F as the main model

In a simple architecture, PWM-1F is the primary model behind your agent:

  1. The agent receives a user request about a project or portfolio.
  2. PWM-1F decides which tools to call to fetch metric histories and any extra context.
  3. Tools return data; PWM-1F reads the histories, reasons about project behaviour and composes the response.
  4. The agent layer handles UX and channel integration; PWM-1F does both orchestration and project reasoning.

This pattern is the default setup and works well when most of the agent’s work is project-centric.

ai for project management architecture
ai for projects in mult-model architecture

2. Multi-model / MoE: PWM-1F as the project specialist

In a multi-model setup, you combine a general foundation model with PWM-1F:

  1. A general model handles broad conversation, routing and non-project tasks.
  2. When the request requires project or portfolio reasoning, the agent:
    • Calls tools to fetch metric histories and relevant context, or
    • Prepares a structured payload for PWM-1F.
  3. PWM-1F reads the histories and returns project-specific analysis.
  4. The general model incorporates that analysis into a broader answer, plan or workflow.

This pattern works well when project questions are one domain among many that a single assistant or copilot must support.

3. Hybrid approaches

Some environments combine the two patterns:

  • Use PWM-1F as the sole model for project-focused agents (e.g. portfolio risk triage),
  • And as a specialist model inside broader assistants that also rely on a general LLM.

And as a specialist model inside broader assistants that also rely on a general LLM.

In all cases, the core idea is the same: your agents bring tools, UX and routing; PWM-1F provides the project-domain reasoning over metric trajectories.

Deployment on Vertex AI

Per-tenant deployments on Vertex AI

PWM-1F is currently delivered as a supervised fine-tuned Gemini model on Google Cloud Vertex AI. Each customer receives:

  • A model deployment in their own GCP project and billing account
  • Integration patterns for calling the model from agents and tools
  • Guidance on prompt design and analysis framing for project scenarios

Customer data flows into their own Vertex environment. PlanVector does not use that data to train the base model.

Integration resources

What you receive as a model customer

  • Metric dictionary and TSV specification
  • Example prompts for core tasks such as project status, variance explanation and early warning
  • Reference patterns for using PWM-1F inside agent frameworks
  • Support during evaluation and rollout
planvector

Building an agent for the project domain?

If you are building agents that need to understand project and portfolio behavior, let us show you how PlanVector can sit at the center of that stack.

Talk to us