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. 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 Resume MCP Registry npm Hugging Face LinkedIn X


Kevin Ruschman

AI Infrastructure Engineer | Agent Systems | Reliability & Verification: "I build the infrastructure that makes AI coding agents reliable."
Target Roles: AI Infrastructure Engineer · Agent Systems Engineer · AI Reliability & Evaluation Engineer · Developer Infrastructure Engineer (AI Agents) · Open to individual-contributor roles, including senior scope depending on team needs · Los Angeles, CA / Remote / Hybrid · Portfolio (krusch.dev) ↗ · Resume ↗

I build the open-source infrastructure that makes AI coding agents reliable: persistent agent memory, deterministic static/AST code verification, crash-recoverable sandboxed execution, and cost-disciplined multi-GPU inference routing. Rather than treating language models as self-policing control planes, I run them as stateless candidate generators bounded by PostgreSQL write-ahead state, <32ms journaled multi-file apply protocols + POSIX atomic per-file renameSync, <15µs pure-CPU Stage-0 cascade dispatch, <1ms static AST checks, and isolated sandboxed test gates, without claiming that deterministic gates guarantee a model's plan is semantically right.

Lab Provenance: Independent R&D lab and open-source engineering research. Projects, evaluations, and reference architectures are developed and tested in my own lab unless otherwise indicated.


🛠️ Open-Source Engineering & Failure-Mode Design

# Run the <1s zero-cloud walkthrough, then initialize the lean 19-tool coding-agent profile (or --profile memory-only / full)
npx krusch-stack demo && npx krusch-stack init --ide all --defer-tools && npx krusch-stack doctor
Watch krusch-context-mcp Narrated Walkthrough (01:31) ▶ krusch-context-mcp v2.0 (01:31)
5-verb SQLite/PG memory & SHA-256 span provenance
Watch @krusch/toolkit: The Golden Bricks Walkthrough (02:26) ▶ @krusch/toolkit v1.8.1 (02:26)
40 runtime Golden Bricks, Brick Exchange & 16 <1ms AST guards
Watch KruschStack Product Tour (01:40) ▶ krusch-stack v1.3.1 (01:40)
Unified 10-pillar, 39-tool stack, 4 profiles & 8/8 MCP Registry
Watch Hermes Parallax Narrated Showcase (02:05) ▶ Hermes Parallax v0.1.0 (02:05)
6-pole epistemic framing, blind AST counterfactuals & DPO forge

1. krusch-context-mcp v2.0.2: Persistent Agent Working Memory & Span Provenance (▶ Video 01:31)

  • Architecture & Adoption (83★ · 6 forks · 355 monthly npm downloads · 155/155 tests): 5-verb Model Context Protocol (MCP) memory engine (retrieve, remember, revise, nudge, health) backed by zero-Docker node:sqlite WAL + external-content Porter FTS5 (or optional PostgreSQL 16 + pgvector HNSW for shared multi-worker state). Enforces SHA-256 line-span provenance (flagging STALE_DRIFT when anchored source lines mutate so stale rules never silently steer edits), zero-LLM git blame -w -M + co-change + causal commit-rationale alignment guards, and $0.00 zero-LLM Raw-RAG (256/32-token sliding conversation windows + 4-Way Hybrid RRF).
  • Scope & Limitations: Single-machine workflows run on built-in node:sqlite (npx krusch-context-mcp init) with zero external services. Zero-LLM Raw-RAG eliminates write-time LLM summarization cost and ingestion hallucination, leaving a 1.67% implicit-synthesis miss floor (10/600 DolphinBench tasks unresolved across Oracle@28) when cross-session constraints share neither lexical tokens nor dense semantic neighborhood at index time.
  • Supporting Benchmarks (Self-Run in Lab): 90.26% accuracy & 95.77% dia_id recall across all 1,540/1,540 Official Snap LoCoMo (ACL 2024) QA probes at $0.00 ingestion (89.68% argon_tight @ 6.73x compression, 92.66% Oracle@2); 81.50% (489/600 @ $1.35, model-matched GPT 5.6 Luna + Grounded, 71.1x cheaper than Mem0), 91.17% (547/600 @ $24.54, Gemini 4 Argon Low, uploaded dolphinbench.ai submission UUID c047b063, status: "queued" awaiting leaderboard acceptance), 90.67% (544/600 @ $30.54, self-run Argon Low + Grounded v0.3, 90.50% on Riley), and 57.33% (344/600 @ $0.00, Wave-6 krush420/gemma4-31b-sre-lora + Grounded; 54.17% stock UUID 288671ad queued) across 28 DolphinBench runs (16,800 traces); 93.00% 10M-token IE recall across 2,000 Official ICLR 2026 BEAM probes.

2. @krusch/toolkit v1.8.1 (JS/TS) & krusch_toolkit v1.6.0 (Python): Certified Runtime Bricks, Collaborative Brick Exchange & 4-Axis Ontology (▶ Video 02:26 · ▶ Ontology 01:38)

  • Architecture & Test Rigor (40 runtime bricks · 65+ .d.ts subpaths · 191/191 Node & 201/201 pytest tests · 0/251 Zero-Leak checks): Pre-audited infrastructure primitives replacing LLM-regenerated boilerplate: connect-time DNS-pinned SSRF guards (safeFetch), bounded HTTP 413 stream parsers, atomic Idempotency-Key locks (409/422), hash-chained SHA-256/SQLite audit-log ledgers, zero-node:* rbac/IDOR guards, HMAC keyset pagination, zero-LLM diff/co-change/causal-commit alignment (@krusch/toolkit/git-alignment), persistent Wayland/CDP browser automation & <2K-tok UI digests (@krusch/toolkit/easi), Continual RSI via Self-Improvement Memory (@krusch/toolkit/modular-rsi, arXiv:2611.18856 counterfactual outcome tracking & cross-model/cross-benchmark transfer), and an Ed25519-sealed multi-engineer & multi-agent Validated Brick Exchange (@krusch/toolkit/brick-exchange) combining Node 22 --permission fault-injection sandboxes, scoped cross-team sharing grants (adapting MIT-licensed 0xANUSHKA/common-ground-8090), a 4-axis cross-stack HydraDB OpenCypher ontology (SoftwareStack, DataPort, ChannelEndpoint, MoEVerificationTier, $0 hexagonal synthesizeBrickBinding), .brick-lock.json drift locks, and 2-hop AST fan-out planning (planAgentFanOut, verifyTwoHopBlastRadius). Paired with krusch-lego's 16 <1ms static/AST vulnerability & boilerplate detectors and 7 zero-LLM project blueprints, validated externally on OWASP/NodeGoat, node-express-realworld-example-app, and stripe-samples/accept-a-payment (plus 104/104 lab repository compliance).
  • Scope & Limitations: The 16 <1ms static/AST detectors enforce structural security and reliability boundaries (raw fetch SSRF, unguarded multipart uploads, yaml.load alias bombs, missing idempotency/RBAC wrappers), not domain-specific business logic or multi-service state invariants, which still require integration suites and independent code review.
  • Supporting Benchmarks: Lifts autonomous open-weights brick adoption from 25% → 100% under pre-commit MCP diff gating and 2-stage QLoRA + DPO calibration (Architecture Paper ↗).

3. krusch-coding-harness + grounded-harness v0.2.1 (& Unified krusch-stack v1.3.1): Sandboxed Staging & Crash-Recoverable Apply (▶ 20s Live Execution · ▶ Stack Tour 01:40)

  • Architecture & Crash Recovery (118/118 + 15/15 harness tests · 91/91 stack tests across 13 suites · 8/8 Official MCP Registry dev.krusch/* servers · 11 npm packages): Stages agent edits inside a PostgreSQL write-ahead apply journal (<32ms), verifies them inside network-disabled Bubblewrap (bwrap --unshare-net) sandboxes, and swaps files onto disk via POSIX atomic renameSync/fsync + pre-commit SHA-256 file-drift guards, verified under a real SIGKILL-mid-apply crash test (scripts/demo-full-harness-loop.mjs: stage flawed patch -> fail sandbox test & reject -> repair & verify -> SIGKILL mid-apply & cold-worker recoverInFlightApplies() -> clean commit). Ships across 10 pillars (21 SQLite + PostgreSQL DDL tables, 39 core + 23 modular extension tools) with progressive onboarding profiles (memory-only 5 tools, coding-agent 19 tools, core 27 tools, full 39 tools), Ed25519 domain-verified Official MCP Registry entries, npx krusch-stack demo, npx krusch-stack explain decision observability, and deterministic clean-room reproduction (npm run bench:repro).
  • Scope & Limitations: For a single interactive developer session, an isolated git worktree and sandboxed test run (the zero-Docker SQLite default in krusch-stack) suffices; PostgreSQL write-ahead journaling targets unattended daemons and multi-worker queues. Each file swap is atomic, whereas a multi-file batch is crash-recoverable rather than atomically visible across files, and a sub-millisecond TOCTOU window remains if an uncoordinated host process mutates the same checkout outside container/worktree isolation.
  • Supporting Benchmarks (Self-Run in Lab): 51.56% on Aider Polyglot (225 tasks) up from 19.56% on the same open-weights model (driven by isolated sandbox test execution, compiler/test error feedback, and bounded repair retries fixing 72 tasks); 0/949 hijacked (0.00% ASR across 1,046 episodes) on ETH SpyLab AgentDojo v1.2.2 (a benchmark-specific finding under active policy floors, bwrap sandboxes, and taint guards, not prompt-injection immunity); 90.67% (544/600 @ $30.54, Argon Low + Grounded v0.3) on DolphinBench; and -95.2% (42 → 2) duplicate MCP calls.

Companion Open-Source Systems

System Engineering Focus Failure-Mode Design & Empirical Evaluation
krusch-cascade-router <15µs Inference Cascade & Circuit Breakers Pure-CPU Stage-0 feature dispatch + HALF_OPEN tier circuit breakers & live SSE logprob aborts (-48.0% 120B TTFT, 1,442x faster Polyglot routing). Maintainer-reviewed & merged upstream on RouteWorks/RouterArena PR #169 (main @ 844e93b; 0.7263 github-actions[bot] CI score across 8,400/8,400 live receipts, leaderboard row queued for next batch ranking refresh) (74.44% acc @ $0.38/1K, 91.43% rob., 26.4x cheaper than GPT-5); clean-room self-run evaluations on Martian RouterBench (36,497 Qs, 0.7488/0.7695 AIQ) & ACL 2026 LLMRouterBench (-83.95% cost).
krusch-git, krusch-fs-mcp & HydraScope AST Blast-Radius, OS Memory & HydraDB Graphs 2-hop AST call-graph blast-radius checks (v1.2.4), SHA-256 page-anchored OS config memory, and self-hosted HydraDB v0.4.0 bi-temporal OpenCypher Metric-to-Surface Dependency Graph + 4-Axis Brick Ontology (v0.1.1, 25/25 tests). 56.40% 10-task Macro NDCG@10 in clean-room ACL 2025 CoIR evaluation (2.39M docs / 162,216 Qs, > Voyage-Code-002 56.26% 512-tok baseline); beats CodeBERT & UniXcoder across 47,998 RepoBench-R queries.
Jean SRE & @krush420 Hermetic SRE Daemon & QLoRA Distillation Containerized repository audit daemon (60/60 tests) + Rule-of-3 scouting; published 18,404 Zero-Leak SFT + DPO rows across 10 splits training gemma4-31b-sre-lora (57.33% @ $0.00, highest open-weights score in self-run 28-run evaluation), 120B Teacher, 30B Student (43.00%), & 8B Edge (15.7x lift).
krusch-law & krusch-nexus Evidentiary Citation Spines & DBOS Workflows Layout-true Poppler TSV document spines (100% held-out citation accuracy) + two-pass proposition entailment enforcing a 0.00% False Support Rate; 26.74% Macro F1 (k=1) & 44.72% Recall (k=2) across 6,889 Stanford LegalBench-RAG queries (714 contracts).
Hermes-Parallax Epistemic Triangulation & Blind AST Counterfactuals Zero-dependency Nous Hermes Agent plugin (Catalog PR #153023, 10/10 pytest, 15/15 validator checks): 6 symmetrical US/global framing poles with Mandarin tokenization + verbatim quote verification, blind AST security entity-scrubbing (168 runs across 7 models), and O_NOFOLLOW poison-quarantined DPO curation.

🤝 Collaborating in External Codebases: Maintainer-Merged Upstream PRs

Building a self-directed R&D lab is different from modifying someone else's codebase under external code review. Here are external Rust, TypeScript, and Python repositories where I diagnosed bugs, followed upstream conventions, and passed maintainer review:

  • hydra-db/selfgraph (9 Merged PRs, Rust & TypeScript): Temporal graph engine fixes covering schema-lineage subtype expansion (#157), endpoint canonicalization (#215), entity-resolution case folding (#180), epochMicros contradiction ordering (#167), plus #149, #156, #179, #208, and #243.
  • RouteWorks/RouterArena (Merged into main @ 844e93b): Multi-tier LLM routing benchmark where maintainer @yl231 audited all 8,400 live OpenRouter predictions and github-actions[bot] CI Run #780 (0.7263 score, 0 abnormal) before merging PR #169.
  • julianosig/easi.run (7 Merged PRs: #3, #4, #5, #7, #8, #9, #10, plus Issue #2, Rust Wayland/CDP): Shipped navigation/editing key delivery, desktop-scoped key & target.pid display routing, UTF-16 .toWellFormed() surrogate-pair protection in inspect_page, MAX_SESSIONS frame attachment bounds, desktop/window display scroll routing, and restarted-controller unheld session cleanup across protocol, display, browser-launch, and admin crates.
  • NousResearch/hermes-agent (Merged PR #153023), hashgraph-online/hol-guard (5 Merged MCP Trust-Binding PRs), & harness-hub/harness-lens#21 (Open HF Space PR): Passed 15/15 catalog validation and plugin_guard static checks to merge the zero-dependency Hermes Parallax epistemic counterfactual plugin, merged 5 MCP server trust-binding specs (#5270, #5290, #5291, #5295, #5296), and submitted verified source references (44/44 + 38/38 checks) for krusch-coding-harness and krusch-grounded-harness to harness-lens.
  • Time-Boxed External Hackathon Sprints (HydraScope & DEADBOLT): Shipped HydraScope v0.1.1 (Live Demo · Write-Up, 25/25 tests) for the HydraDB Hackathon (self-hosted hydradb:0.4.0 OpenCypher + @hydradb/selfgraph agent context layer compressing 64 MCP tool schemas by -94.2%, 4-axis cross-stack brick ontology, and bi-temporal metric-to-surface dependency propagation, surfacing the 9 merged hydra-db/selfgraph PRs above) and DEADBOLT (Architecture Paper) in-person under a 20-hour code freeze at EVOLVE (8090 AI Hackathon, Redwood City), integrating 8090's Work Order APIs and HydraDB v0.4.0.

🔧 Engineering Background

Before focusing full-time on AI infrastructure in late 2025, I spent twelve years (2013–2025) working hands-on in physical systems and operations: executing NEC-compliant residential electrical installations/remodels (panel upgrades, kitchen/bath, recording studio) and deploying festival generator and stage wiring across New England (Atomic Audio & Visual, North Beast, MASS MoCA, plus Stratton Trail Crew) out of Jamaica, Vermont (2020–2025); mentoring local teens on the Mobile Bike Repair team at the Wisconsin Bike Fed in Milwaukee (2019–2020); and owning and operating Local Coffee in downtown Aspen after helping launch the Gorsuch Ski Café at the base of Ajax Mountain and Venue in downtown Fairbanks, Alaska (2013–2019, building on Troop 883 Eagle Scout leadership and cycling from San Diego to Chiapas at age 19). Working where miswired circuits, overloaded generators, and unhedged operating costs fail immediately shaped my engineering instincts around failure modes and operational accountability. That background is why I govern language models with deterministic engineering gates (PostgreSQL write-ahead state, <15µs pure-CPU dispatch, and <1ms static AST checks) rather than relying on unverified LLM behavior.


📊 Benchmark Evidence & Publication Status

Systems & Performance Benchmarks

Evaluated across 335,000+ traces and 14 external benchmark suites in my lab, with published SHA-256 audit logs and reproducible evaluation artifacts (95/95 SHA-256 clean-room verification gates across 49 Hugging Face audit logs in 64 files + 15/15 adyen/DABstep gates + 5/5 Official Snap LoCoMo gates, enforcing runtime JS Proxy oracle-leakage traps, 0o444 read-only SQLite freezes, and loopback air-gaps on open-weights runs). Running a public benchmark in my own lab is distinct from third-party reproduction or leaderboard acceptance, so every score is labeled with its exact status:

  1. Maintainer-Reviewed & Merged Upstream (github-actions[bot] CI-Verified): RouteWorks/RouterArena PR #169 (0.7263 CI Arena Score across 8,400/8,400 live OpenRouter receipts + 420 robustness noise perturbations; merged upstream by maintainer @yl231 into main @ 844e93b; leaderboard row queued for next batch ranking refresh).
  2. Uploaded dolphinbench.ai Submissions (status: "queued" Awaiting Leaderboard Acceptance): 91.17% (547/600 @ $24.54, Gemini 4 Argon Low, UUID c047b063, exceeding the current 70.67% published leaderboard entry in my 600/600 evaluation), 89.17% (535/600 @ $24.57, Argon Medium, UUID 29389157), and 54.17% (325/600 @ $0.00, Google Gemma 4 31B stock, UUID 288671ad), among 8 uploaded .zip archives passing submission.validate. Until accepted onto the leaderboard, read these as self-run results.
  3. Self-Run Clean-Room Evaluations (Published SHA-256 Audit Logs): 90.67% (544/600 @ $30.54, Argon Low + Grounded v0.3, 90.50% on Riley), 57.33% (344/600 @ $0.00, Wave-6 krush420/gemma4-31b-sre-lora + Grounded), Official Snap LoCoMo (90.26% accuracy / 95.77% dia_id recall), ETH SpyLab AgentDojo v1.2.2 (0/949 hijacked across 1,046 episodes under active policy/sandbox defenses), adyen/DABstep (9/10 tasks, small sample), ACL 2025 CoIR, ICLR 2026 BEAM, BFCL v3 Live, Aider Polyglot (51.56%), BIRD-SQL, RepoBench-R, LegalBench-RAG, RouterBench, and LLMRouterBench: Full Methodology & Honest Gaps (BENCHMARKS.md) ↗ · Raw SHA-256 Logs on Hugging Face ↗.
  4. Third-Party Reproduction: None Yet (Invitation Open): No outside party has re-run these benchmarks yet. Scripts, frozen read-only memory stores, and all 49 raw log sets (64 files) are public on Hugging Face so anyone can inspect or reproduce them; reproduction attempts are welcome as GitHub issues.

How I Keep Self-Run Benchmarks Clean (5 Code-Enforced Anti-Leakage Controls): (1) Runtime Answer-Key Traps (JS Proxy & Python BlindMappingGuard) abort immediately with FATAL_ORACLE_LEAKAGE_VIOLATION if ingestion code ever touches ground-truth fields (qa, ideal_response, probing_questions, source_chat_ids, .gold_sql, or session_summary); (2) Read-Only 0o444 Frozen Databases index conversations once at $0.00, hash memory_manifest.json, and lock chmod 0444 on disk before any evaluation question is read; (3) Zero Dataset-Specific Tuning audits src/*.js for zero benchmark proper nouns or rubric word lists (Snap LoCoMo ran the byte-identical SHA-256 engine built before downloading locomo10.json); (4) Apples-to-Apples Prompts & Disclosed Token Budgets run Mem0's verbatim ANSWER_PROMPT_ZEP + ACCURACY_PROMPT (T=0) across both 4,465-token (89.68%) and 9,983-token (90.26%) LoCoMo budgets (95.77% dia_id recall scored separately) and hold GPT 5.6 Luna (T=0) fixed on DolphinBench (81.50% @ $1.35 vs. Mem0 70.67% @ $96.21); and (5) 3-Second Offline Verifiers (95/95 Gates) let anyone run npm run bench:beam:verify in krusch-context-mcp with zero API keys or Docker to re-verify source digests, canary traps, and per-question recall tallies from raw logs.


🧰 Core Engineering Competencies & Technical Stack

  • Core Engineering Competencies: Software Architecture & Systems Engineering · Distributed & Crash-Recoverable State · AI Agent Orchestration & MCP · Application & Sandbox Security · Automated Testing & Verification · Inference & Latency Optimization · Open-Source Collaboration (Rust, TypeScript, Python)

  • Languages & Runtimes: JavaScript / TypeScript (Node.js 22+ ESM, React, Vite, Express, Cloudflare Workers, oxlint) · Python 3.12 (FastAPI, SQLAlchemy 2.0, Pydantic v2, PyTorch, asyncio) · SQL

  • State, Graph & Search: PostgreSQL 16 (pgvector HNSW, tsvector GIN, Hybrid RRF) · node:sqlite WAL + external-content Porter FTS5 · snowflake-arctic-embed2 (8,192-ctx, 1,024-dim, asymmetric query: calibration)

  • Infrastructure & LLMOps: Multi-Node Bare-Metal Linux (10GbE, systemd, Docker, Bubblewrap sandboxes) · Ollama / vLLM (q8_0 KV) · 96GB RTX PRO 6000 Blackwell QLoRA + DPO · GitHub Actions CI


📖 Selected Architecture Writing (All 30 Papers on krusch.dev →)


Engineered by Kevin Ruschman (@kruschdev) · Los Angeles, CA · krusch.dev

Pinned Loading

  1. krusch-stack krusch-stack Public

    Local SQLite control plane for AI coding agents: persistent memory, Turn-0 grounding, spin-loop guards, and 2PC file safety.

    JavaScript 3

  2. toolkit toolkit Public

    39 pre-tested, security-audited building blocks for AI apps plus a sub-millisecond scanner for common AI-generated code bugs.

    JavaScript

  3. krusch-context-mcp krusch-context-mcp Public

    Persistent working memory for AI coding agents: remembers project rules with SHA-256 span drift at bash AI cost (SQLite/PostgreSQL). 90.26% on Snap LoCoMo.

    JavaScript 83 6

  4. 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

  5. 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…

    JavaScript

  6. 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