Palantir in 10 Minutes: Foundry, Gotham, AIP and the Ontology
Explore how Palantir’s Foundry, Gotham and AIP differ and how the Ontology turns raw records into operational decisions. Use a maintenance example to distinguish object types, objects, links, actions and applications. Trace how a validated action changes operational state, and recognize why an AI suggestion alone does not validate or execute a decision. Review what the FDSE role expects you to know about Palantir’s products, and see how this course builds toward a full mock interview loop.
We'll cover the following...
- What does assigning repair R42 require?
- Where do Foundry, Gotham and AIP fit?
- How do three records become an assignment workflow?
- What does a schema comparison leave out?
- Can you make the decision without a model?
- What should you already know before preparing?
- How does preparation build toward a mock loop?
Palantir provides software platforms for connecting organizational data to decisions and operational work. Consider a maintenance coordinator who needs to assign an open repair: Equipment records, repair requests and technician availability must come together before anyone can make a useful assignment. Following that decision from source records to a completed action makes Palantir’s product vocabulary concrete,and gives you a grounded starting point for Forward Deployed Software Engineer interviews.
What does assigning repair R42 require?
Suppose pump P17 needs repair R42. The coordinator needs to know which equipment the repair concerns, which technicians are available and which are qualified for pumps. A useful application brings those facts together and lets the coordinator assign the work under defined rules.
Palantir’s platforms support integrating data and building operational applications, including AI-enabled workflows. Foundry supplies data and workflow-building capabilities, while AIP supplies capabilities for using generative AI within those workflows.
That connection to actual work matters for this role. Forward deployed engineering embeds engineers in customer environments, working alongside users to build and ship solutions to operational challenges. FDSE means Forward Deployed Software Engineer.
Your starting question is therefore practical: What must the coordinator be able to see, decide and change?
Where do Foundry, Gotham and AIP fit?
These names describe related offerings with different focuses:
Foundry: Provides capabilities for data management, logic authoring, Ontology development, analytics and workflow development. In the maintenance case, these capabilities support bringing records together and building the assignment workflow.
Gotham: Palantir’s original platform, focused on defense, intelligence and public-sector users who need to integrate data and support operational decisions. The split from Foundry is one of emphasis, not an absolute government-versus-commercial boundary.
AIP: Provides secure large language model connectivity, agent development tools, AI-enabled applications and evaluation capabilities. In the maintenance case, an AI assistant might suggest a technician or summarize a repair request.
AIP is not another name for Foundry or the Ontology. Likewise, the Ontology is not a fourth equivalent platform in this comparison: It supplies the connected objects, relationships and actions that applications use.
These offerings can cooperate. The maintenance workflow remains useful even when no model is involved; adding model assistance introduces another capability whose output the workflow must handle correctly.
How do three records become an assignment workflow?
Start with these source records:
{"equipment": {"id": "P17", "type": "pump"},"repair": {"id": "R42", "equipment_id": "P17", "status": "open", "assignee_id": null},"technician": {"id": "T8", "available": true, "qualification": "pump"}}
The repair’s equipment_id identifies P17, while its empty assignee_id shows that nobody has been assigned yet.
An object type defines a real-world entity or event; an object is one instance of that type. Here, Pump, Repair and Technician are object types, while P17, R42 and T8 are their individual objects. Ontology objects can connect to backing data or be created through configured actions.
A link type defines a relationship between object types; a link is one instance of that relationship. The link type “Repair concerns Pump” describes the relationship’s structure. “R42 concerns P17” is the particular link in this case. These concepts connect to actual organizational data, allowing applications to work with the equipment and repairs the organization tracks.
This is a conceptual flow, rather than a deployment architecture. Foundry supplies capabilities across the data-to-workflow process; the diagram separates the concepts so you can explain what each contributes. A repair-queue application presents the connected information and an assignment control to the coordinator.
What does a schema comparison leave out?
If you already write code, comparing object types to table definitions is a useful starting point. A Repair object type describes structure, and R42 resembles a particular row. But that comparison alone leaves the operational behavior unexplained: Which equipment does this repair concern, what action can a user take, and what must hold before that action succeeds?
In this case, distinguish five things:
Object types: Pump, Repair and Technician define the kinds of entities represented.
Objects: P17, R42 and T8 identify particular entities.
Relationship: R42 concerns P17, connecting the repair to its equipment.
Action: Assign repair changes the repair’s assignment under the example’s rules.
Application: The coordinator’s repair queue exposes the information and action through a user interface.
Ontology Manager supports defining object types and action types, connecting data to the Ontology and investigating whether data is updating in user applications.
A repair queue can display the same objects that other applications use. That makes the distinction useful: The application is the user-facing experience; the objects, links and configured action describe what that experience works with.
For this example, assignment requires an available technician qualified for the equipment type. The coordinator selects Assign repair with R42 and T8:
R42 concerns P17, whose equipment type is
pump.T8 is available and has the
pumpqualification, so both checks pass.The action changes R42’s status to
assignedand its assignee toT8.R42 still concerns P17; assigning a technician does not change which equipment needs repair.
The operational result comes from applying a validated action to connected objects. This describes the example’s resulting operational state; propagation to an external source system requires its own configured integration behavior.
Can you make the decision without a model?
Add technician T9, who is available but qualified only for electrical work. Suppose a model suggests T9 because it notices availability and overlooks the pump qualification.
Availability satisfies only one part of the rule. An attempted assignment to T9 must be rejected, leaving R42 open with assignee_id = null. A plausible suggestion has not established eligibility or executed an assignment.
The tempting T9 choice treats availability as sufficient, confusing one satisfied condition with the complete rule. Keeping validation separate from recommendation prevents that shortcut from changing operational state. You can make the correct decision from the records alone, with no Palantir account, code or AI model.
What should you already know before preparing?
This course is for candidates who can write code and use basic data structures, whether their evidence comes from projects or professional engineering work. SQL and system design may be new. Prior Palantir platform mastery is not assumed.
The preparation goal is to make your reasoning, implementation, learning, debugging and communication observable in interviews. Product vocabulary helps you explain an operational problem precisely; it does not replace demonstrating those skills or guarantee an offer.
How does preparation build toward a mock loop?
The course uses independent examples, so exercises do not depend on hidden state from a shared project. Four chronological stages connect this orientation to interview practice. Explore each stage’s outcome to see what you will be able to demonstrate.
This coverage is a preparation route, rather than a promise that every candidate receives every round. Public candidate reports describe differing combinations of interviews; confirm your own process with your recruiter.
Review the draft explanation below and identify its mistaken claims. Then give another candidate a short corrected explanation using the supplied records and assignment rule. Connect the products’ capabilities to what the coordinator actually changes.
The products contribute different capabilities:
Foundry: Data management, logic, Ontology, analytics and workflow-building capabilities.
AIP: Secure model connectivity, agent tools, AI-enabled applications and evaluation,not the whole operational system.
Gotham: Defense, intelligence and public-sector operational decision-making, as a focus rather than a strict sector boundary.
Connect those capabilities to the example:
Objects and link: P17, R42, T8 and T9 are object instances. R42 concerns P17 is the repair-to-equipment link.
Application and action: The repair queue is the application; Assign repair is the action.
Eligibility: T8 meets both requirements: availability and pump qualification. T9 fails the qualification check.
Result: A successful assignment sets R42’s status to
assignedandassignee_idtoT8, preserving its equipment link to P17.AI’s limit: Optional AI assistance can suggest a technician, but a suggestion neither validates nor executes the assignment.
The unchanged equipment link matters as much as the changed assignee: a correct action changes only the intended operational facts. Next, consider what engineering work an FDSE takes ownership of to deliver such a workflow.