Skip to content
View kruschdev's full-sized avatar
💭
Got RAG?
💭
Got RAG?

Block or report kruschdev

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
kruschdev/README.md
Kevin Ruschman — @kruschdev



Website Companion MCP RouterArena PostgreSQL LinkedIn X


Sovereign Personal AI · Local-First Agent Infrastructure · Sub-15µs Cascade Routing


💡 Engineering Premise

Language models are probabilistic samplers; modifying your codebase or managing confidential client knowledge requires a deterministic transaction manager.

I build sovereign, local-first agent infrastructure so AI coding agents can plan, navigate, and verify multi-step tasks without mutating your physical disk until verified. State lives in PostgreSQL, models remain completely swappable, and operations enforce strict transactional invariants.


🛠️ The 3-Tier Sovereign Agent Ecosystem

Standardized across Antigravity, Claude Code, Cursor, and Windsurf:

3-Tier Sovereign Agent MCP Ecosystem

Tier Subsystem Authoritative Responsibility Key Capabilities
Tier 1 krusch-context-mcp Working Memory & Invariants 5-verb sovereign memory engine (retrieve, remember, revise, nudge, health). Eliminates agent amnesia with ACID decision logs, lineage tracking, and pre-commit invariant gates.
Tier 2 krusch-git AST Code Graph & Exploration Full Git DAG in PostgreSQL, AST symbol extraction, caller-callee call graphs via recursive relational CTEs, 40-char SHA pinning, and exponential temporal recency decay.
Tier 3 krusch (Harness) Staged Execution & Verification PostgreSQL state machine, invariant gate before mutation, virtual unified diff staging, and two-phase commit (2PC) disk application. Zero uninspected disk writes.
Bridge @krusch/polygres-connector Zero-Friction Cloud Substrate Cloud connector marrying KruschContext with pgGraph and Wondersearch with 40-char SHA pinning, cross-substrate joins, and strict ABA Model Rule 1.6 air-gap protection.

🚀 Flagship Systems & Open-Source Projects

🛡️ krusch-companion · Sovereign Personal AI & Daily Companion

A 100% private, local-first alternative to cloud-tethered personal assistants (like OpenAI Dot or xAI Grok Companion). Triages raw voice thought streams into actionable 5-tier priority boards and maintains persistent working memory on local SQLite.

  • Zero Cloud Egress: 100% local SQLite WAL storage; zero prompt training or data harvesting.
  • Microsecond Routing: Employs krusch-cascade-router (<15µs CPU L0 fast-path) and local Ollama edge models for $0.00 daily interaction costs.
  • Zero Data Retention Cloud Mode: Optional OpenRouter integration injecting provider: { data_collection: "deny" } for zero-GPU setups while keeping memory local.
  • Installable PWA: React 19 + Vite luxury dark glassmorphism dashboard with ambient focus timer and daily coaching reflections.

🧠 krusch-context-mcp · Persistent Working Memory & Invariant Steering

A lightweight, 5-tool Model Context Protocol server giving AI coding agents (Claude Code, Cursor, Windsurf, Antigravity) persistent memory and rule enforcement across sessions.

  • 5 Core Verbs: retrieve, remember, revise, nudge, health constrained to <350 prompt tokens.
  • Temporal Fact Superseding: Invalidation engine with lineage links that retires obsolete rules and prevents contradictory prompt context.
  • Pre-Commit Invariant Gate: Audits code diffs against project rules before disk mutation to prevent silent regressions.
  • SQLite-First: Zero-Docker setup built on node:sqlite with optional PostgreSQL dual-engine sync.

⚡ krusch-cascade-router · Sub-15µs CPU Routing

Dual-stage microsecond cascade router designed for CPU-only execution. Evaluates semantic difficulty, token length, query structure, and sensitivity to route requests across free, cheap, and frontier models.

  • Latency: Sub-15µs L1 heuristic gating; sub-8ms L2 dense vector centroid scoring.
  • Speculative Hedging: Automatically dispatches hedge requests when fast models return ambiguous outputs.
  • Cost Reduction: Slashes aggregate inference spend by 7–9% while beating baseline Sonnet 3.5 quality scores.

Cloud connector bridging local sovereign agent memory and codebase graphs with Polygres Cloud and Wondersearch.

  • Cross-Substrate Marriage: getSymbolContext() resolves AST definitions, recursive caller CTEs, and active steering invariants in a single database round-trip.
  • Refactor Safety Gate: Automatically flags requiresCallerAudit = true whenever a symbol with inbound callers is modified or renamed.
  • Air-Gap Boundary: Enforces fail-closed isolation for privileged litigation matters and confidential commercial contracts.

⚖️ krusch-law

100% air-gapped sovereign legal intelligence platform.

  • Zero Hallucination Grounding: Enforces a 0.00% False Support Rate across statutory tenant protection codes and appellate cases.
  • Statutory Precedence DAG: Traverses jurisdictional overrides and local ordinances with strict citation spine coordinates.
  • Air-Gap Compliance: Runs strictly on local PostgreSQL 16 with realpath denial blocking cloud egress under ABA Model Rule 1.6.

Universal high-throughput document extraction and dual-provider RAG substrate.

  • Deterministic TOC Parser: Extracts hierarchical document outlines in <5ms.
  • 4-Stage Retrieval Ablation: Vector-only, FTS-only, hybrid, and RRF + local cross-encoder reranking.

💻 Tech Stack & Sovereign Philosophy

  Languages        : TypeScript (ESM) · Python 3.12 · Rust · SQL · C/C++
  Databases        : PostgreSQL 16 · pgvector · SQLite (WAL) · DuckDB
  Agent Protocol   : Model Context Protocol (MCP) · JSON-RPC · SSE Streams
  Code Analysis    : Tree-Sitter AST · Relational Call Graphs (Recursive CTEs)
  Infrastructure   : Local Homelab Fleet · Linux (Ubuntu/Pop!_OS) · Docker
  • PostgreSQL-First Proxy: Operational data, AST symbols, and agent episodic memories persist in relational PostgreSQL with strict relational foreign keys and indexes.
  • Fail-Closed Air-Gaps: Cloud egress is opt-in (ALLOW_CLOUD=1). Private files, counterparty terms, and litigation matters remain strictly on-premises.
  • Deterministic Over Probabilistic: Prompt chains are constrained by state machines, invariant verification gates, and atomic rollback journals.

📊 Verified Systems & Performance Benchmarks

Systems & Performance Benchmarks

GitHub Streak Stats

📖 Selected Articles & Deep Dives


Engineered by Kevin Ruschman (@kruschdev) · krusch.dev

Pinned Loading

  1. krusch-companion krusch-companion Public

    Sovereign, local-first personal AI assistant & daily companion (like Grokbot, OpenAI Dot, or Muse). Captures ambient voice thought streams into prioritized action boards with durable local memory, …

    JavaScript

  2. krusch-context-mcp krusch-context-mcp Public

    Persistent working memory and invariant steering MCP server for AI coding agents (Claude Code, Cursor, Windsurf). Captures architectural decisions, enforces project rules with pre-commit audits, an…

    JavaScript 71 6

  3. krusch-git krusch-git Public

    Git DAG indexing, AST symbol graphs, and hybrid semantic code search MCP server for AI coding agents (Claude Code, Cursor, Windsurf). Maps repositories into PostgreSQL + pgvector with recursive cal…

    JavaScript

  4. krusch-nexus krusch-nexus Public

    Air-gapped multi-format document ingestion, local OCR fallback, structural legal/business chunking, and PostgreSQL/pgvector indexing for the Krusch Ecosystem.

    Python 1 1

  5. krusch-law krusch-law Public

    Private, air-gapped legal RAG & ordinance intelligence engine for on-premise case analysis with zero cloud data leakage.

    Python

  6. krusch-cascade-router krusch-cascade-router Public

    Dual-stage cascade router for LLMs and agent swarms. Intercepts code, SQL, and math on CPU in <15µs for $0.00 without pre-flight model calls—saving 300–800ms TTFT per turn. Features L2 neural centr…

    Python