I am an AI-native systems engineer based in Los Angeles specializing in autonomous agent harnesses, enterprise context engines, and transactional state stores. While traditional development paradigms struggle with probabilistic model drift and unmediated file writes, my engineering centers on hard systemic invariants: staging agent diffs in PostgreSQL ACID tables, verifying execution inside isolated test sandboxes, and committing changes strictly via two-phase commit (2PC) fsync gates.
As the creator of krusch-context-mcp (a 71★ Model Context Protocol server eliminating agent amnesia across sessions) and krusch (a PostgreSQL-backed coding harness enforcing test verification), I have hands-on experience designing the exact tool-use protocols, memory substrates, and guardrails required by Slalom's Fortune 500 enterprise clients across AWS Bedrock, Azure AI Foundry, and Google Cloud Vertex AI.
A direct, line-by-line mapping between Slalom's West Region job requirements and my shipped production codebases, architectural papers, and open-source tooling.
| Slalom Role Requirement | Kevin's Shipped Architecture & Production Proof | Direct Evidence |
|---|---|---|
|
Model Context Protocol (MCP) Integration:
"Design and implement MCP server integrations for tool use, data access, and cross-system interoperability..."
|
Creator and maintainer of krusch-context-mcp (71★ on GitHub). Engineered a unified 16-tool Model Context Protocol server over stdio JSON-RPC transport with mathematical temporal recency decay (e^-0.01t), declarative steering facts, AST symbol graphs, and multi-agent consensus. Ready to lead Slalom's enterprise MCP practice.
|
71★ GitHub Repo |
|
HITL Guardrails & Sandbox Isolation:
"Build guardrails for safety, compliance... including human-in-the-loop (HITL) approval workflows and sandbox isolation..."
|
Architected krusch and kd-Code. Replaces naive file-writing with a PostgreSQL-backed transactional harness: models are treated as ephemeral workers, proposed edits are staged as SHA-256 diffs in relational tables, test suites execute inside unprivileged Bubblewrap (bwrap) user-namespace sandboxes (read-only root, network isolation), and physical disks are updated strictly via atomic write-ahead apply journals upon human review.
|
Harness Architecture |
|
Multi-Agent Orchestration & Planning:
"Design end-to-end agentic AI architectures including planning loops, memory management, tool integration, and coordination patterns..."
|
Designed multi-agent consensus mechanisms (Company Brain v2) with versioned state graphs, automated conflict resolution, and role-based access lenses. Engineered speculative parallel hedging and multi-specialist dispatch in krusch-cascade-router, orchestrating open and frontier models dynamically.
|
Multi-Agent State |
|
Advanced Enterprise RAG & Vector Systems:
"Vector databases, embedding strategies, chunking optimization, hybrid search, re-ranking, and multi-source data synthesis..."
|
Engineered production PostgreSQL 16 hybrid retrieval stacks fusing tsvector cover-density text search with pgvector HNSW cosine distance via Reciprocal Rank Fusion (RRF). Authored multi-hop dependency graph walks, zero-dependency structural symbol chunking, and deterministic server-side token budget packing.
|
Production pgvector |
|
AI-Assisted Dev Workflows (Claude Code / Cursor):
"Experience with AI-assisted development tools such as Claude Code, Cursor, Kiro... for code generation, debugging, and workflow acceleration..."
|
Daily power user and builder in Claude Code, Antigravity IDE, and Cursor. Built kd-Code as a custom developer control plane featuring side-by-side diff reviews via @pierre/diffs, AST symbol impact graphs, and multi-model hot-swapping between Claude 3.5 Sonnet, Gemini 2.5, and local Ollama runtimes.
|
Developer Tooling |
|
Client Communication & Executive Advisory:
"Explain model behavior to technical and non-technical audiences, leading deep technical presentations and architecture conversations..."
|
Demonstrated ability to translate complex AI systems into high-stakes business proposals and compliance runbooks. Recently authored the dedicated Younessi Law Sovereign AI Proposal, creating a 30-day closed-file pilot framework with 5 empirical gating milestones (OCR confidence, wage-hour math, and CCP § 128.7 grounding). | Enterprise Proposal |
Battle-tested architectures ready to deploy into Slalom's enterprise client engagements across AWS, Azure, and GCP.
Decoupling access governance, working-memory density, and sovereign state for Claude Code, Cursor, and custom agent runtimes without agent amnesia or prompt bloat.
Eliminating hallucination risk and unmediated file clobbering when autonomous coding agents execute complex refactors on enterprise codebases.
Slashing enterprise LLM inference budgets by 60–80% by routing deterministic queries on CPU before escalating to expensive frontier models.
Real, inspectable open-source software backing my engineering credentials.
Production Model Context Protocol server exposing 16 tools across episodic memory, steering nuggets, AST symbol graphs, and routing centroids. Eliminates agent amnesia across sessions.
PostgreSQL-grounded coding harness enforcing diff-staging, sandboxed test verification, single-writer file leases, and atomic 2PC fsync commits. Zero unvetted file writes.
Dual-stage cascade router unifying an L1 CPU syntactic fast-path with L2 neural centroid escalation, speculative hedging, and specialist dispatch across open and frontier models.
A predictable, consultative 4-week delivery framework designed to take enterprise clients from initial architecture discovery to secure, observable production systems.
Assess client workflows, map data boundaries (PII/HIPAA/IP), identify hallucination risk vectors, and establish human-in-the-loop (HITL) approval governance policies.
Stand up dedicated MCP context servers, configure vector embeddings with PostgreSQL 16 + pgvector, and implement temporal recency decay across client knowledge stores.
Deploy isolated verification runners for agent actions, wire in multi-tier cascade routing to optimize token economics, and integrate automated evaluation metrics (task accuracy, tool precision).
Deliver telemetry dashboards, audit logging runbooks, and lead hands-on client engineering workshops—ensuring seamless operational transition and long-term autonomy for the client team.
I am actively interviewing for the Agentic AI Architect / Senior Engineer role in Los Angeles. I would welcome the opportunity to discuss how my hands-on agent infrastructure background can accelerate Slalom's Intelligence Engineering practice.