AI

Palantir

We provide Foundry and AIP implementation and integration for teams using Palantir. We connect data sources, shape operational workflows and integrate the resulting applications with existing systems, with attention to access controls and ongoing ownership.

Agentic Harness Development

We design and run autonomous software-delivery harnesses that turn scoped work into reviewable changes. Each harness brings task dispatch, code context, execution limits and delivery evidence into a coordinated workflow.

• GitHub-issue-driven task dispatch connects each task to its requirements and acceptance criteria.
• AWS Step Functions orchestration coordinates execution stages, retries and escalation paths.
• Amazon Bedrock AgentCore runtime provides the execution environment for delivery agents.
• Strands Agents orchestration with OpenCode as the code-editing engine connects agent decisions to repository changes.
• Tree-sitter repository maps for code context help agents locate relevant definitions and relationships.
• Multi-model routing with cost and time budgets and tier escalation bounds work and routes difficult tasks to the next tier.
• An independent, commit-bound code-review gate checks the exact revision proposed for delivery; no change leaves draft unreviewed.
• CI checks on every commit keep validation tied to the current code rather than an earlier result.
• DynamoDB task state with fenced ownership prevents a stale worker from taking over an active task.
• Self-hosted runners and isolated container builders separate check execution from container image builds.
• A live status dashboard makes progress, review outcomes and blocked work visible to the people overseeing delivery.

Model Safety

We develop governance and safety controls for AI systems: defined authority, evaluation criteria, traceable decisions and explicit intervention points. The forthcoming MP-AIA framework focuses on how these controls fit into day-to-day engineering and operational review.

Read the MP-AIA whitepaper (forthcoming)

Agentic self-healing operations

We build agents that detect, diagnose and remediate operational failures under deterministic guardrails and human approval. Allowed actions, verification steps and escalation conditions are defined in advance; remediation requires human approval, and uncertain cases return to an operator with supporting evidence.