Full-Stack & Geospatial Engineer · Agentic Systems Builder
Founder @ Khamseen Technologies

I build production systems at the intersection of software engineering, geospatial platforms and AI orchestration.

• Agent infrastructure — control planes, runtime abstractions, skills, policy gates, audit trails and autonomous execution.
• Geospatial platforms — ArcGIS Experience Builder, ArcGIS Maps SDK for JavaScript, spatial workflows and API-driven GIS applications.
• Full-stack systems — TypeScript/React frontends, Python/Django backends, PostgreSQL, Redis and containerized infrastructure.
• Reliability — deterministic checks, explicit failure states, independent QA, bounded retries and observable execution.

Agent-native enterprise control plane designed to coordinate humans, autonomous agents, deterministic systems, tools and replaceable execution runtimes.

Khamseen owns the semantics that should not disappear inside an LLM or a vendor runtime:

Agent · Task · Run · Unit · Event · ToolCall · Decision · Gate · Approval · Audit

The architecture deliberately separates the control plane from the execution plumbing. Khamseen keeps authority, policy and business state; interchangeable infrastructure can handle execution, isolation, tools and model access.

Graph ≠ Loop. A graph decides which Units should exist, what actually depends on what, and what can run in parallel. A loop converges one Unit toward correctness:

produce → check → correct → repeat → escalate

Capability ≠ Skill ≠ Tool. Capabilities define what an agent is authorized to do. Skills encode the procedure for doing it. Tools are external actions. The Agent Skills layer is being designed around versioned SKILL.md contracts and progressive loading so agents receive procedural context only when it is needed.

Runtime plumbing stays replaceable. Execution flows through abstractions such as RuntimeAssignment, harnesses, sandboxes and gateways rather than binding agent identity to a model or vendor. Hermes is integrated behind the runtime boundary; OpenClaw is being evaluated through a bounded external-runtime adapter spike. Cursor, Codex, Claude Code and future providers can sit behind the same architectural contract.

Deterministic before LLM. Tests, type checks, schema validation, git diff/scope checks, dependency state, authorization and resource budgets should decide what they can before model judgment is used. Implementation and independent review remain separate execution contexts.

Human authority stays explicit. The objective is not “AI with no humans”; it is moving human involvement toward goals, exceptions and consequential approvals instead of manually babysitting every intermediate step.

_{Building systems where agents can act autonomously without making authority, evidence or failure disappear.}