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.
# 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
▶ krusch-context-mcp v2.0 (01:31)5-verb SQLite/PG memory & SHA-256 span provenance |
▶ @krusch/toolkit v1.8.1 (02:26)40 runtime Golden Bricks, Brick Exchange & 16 <1ms AST guards |
▶ krusch-stack v1.1.0 (01:40)Unified 7-pillar, 39-tool stack & 2PC + Grounded harnesses |
▶ 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★·6forks ·86monthly npm downloads ·155/155tests): 5-verb Model Context Protocol (MCP) memory engine (retrieve,remember,revise,nudge,health) backed by zero-Dockernode:sqliteWAL + external-content PorterFTS5(or optionalPostgreSQL 16 + pgvector HNSWfor shared multi-worker state). Enforces SHA-256 line-span provenance (flaggingSTALE_DRIFTwhen anchored source lines mutate so stale rules never silently steer edits), zero-LLMgit blame -w -M+ co-change + causal commit-rationale alignment guards, and$0.00zero-LLMRaw-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-LLMRaw-RAGeliminates write-time LLM summarization cost and ingestion hallucination, leaving a1.67%implicit-synthesis miss floor (10/600DolphinBenchtasks unresolved acrossOracle@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_idrecall across all1,540/1,540Official SnapLoCoMo(ACL 2024) QA probes at$0.00ingestion (89.68%argon_tight@6.73xcompression,92.66%Oracle@2);81.50%(489/600@$1.35, model-matchedGPT 5.6 Luna + Grounded,71.1xcheaper thanMem0),91.17%(547/600@$24.54,Gemini 4 Argon Low, uploadeddolphinbench.aisubmissionUUID c047b063,status: "queued"awaiting leaderboard acceptance),90.67%(544/600@$30.54, self-runArgon Low + Grounded v0.3,90.50%onRiley), and57.33%(344/600@$0.00, Wave-6krush420/gemma4-31b-sre-lora+Grounded;54.17%stockUUID 288671adqueued) across28DolphinBenchruns (16,800traces);93.00%10M-token IE recall across2,000Official ICLR 2026BEAMprobes.
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 (
40runtime bricks ·65+.d.tssubpaths ·191/191Node &201/201pytesttests): Pre-audited infrastructure primitives replacing LLM-regenerated boilerplate: connect-time DNS-pinned SSRF guards (safeFetch), bounded HTTP413stream parsers, atomicIdempotency-Keylocks (409/422), hash-chained SHA-256/SQLiteaudit-logledgers, zero-node:*rbac/IDOR guards, HMAC keysetpagination, zero-LLM diff/co-change/causal-commit alignment (@krusch/toolkit/git-alignment), and an Ed25519-sealed multi-engineer & multi-agent Validated Brick Exchange (@krusch/toolkit/brick-exchange) combining Node 22--permissionfault-injection sandboxes, scoped cross-team sharing grants (adapting MIT-licensed0xANUSHKA/common-ground-8090), a 4-axis cross-stack HydraDB OpenCypher ontology (SoftwareStack,DataPort,ChannelEndpoint,MoEVerificationTier,$0hexagonalsynthesizeBrickBinding),.brick-lock.jsondrift locks, and 2-hop AST fan-out planning (planAgentFanOut,verifyTwoHopBlastRadius). Paired withkrusch-lego's16<1msstatic/AST vulnerability & boilerplate detectors and7zero-LLM project blueprints, validated externally onOWASP/NodeGoat,node-express-realworld-example-app, andstripe-samples/accept-a-payment(plus104/104lab repository compliance). - Scope & Limitations: The
16<1msstatic/AST detectors enforce structural security and reliability boundaries (rawfetchSSRF, unguarded multipart uploads,yaml.loadalias 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+15/15harness tests ·91/91stack tests): Stages agent edits inside aPostgreSQLwrite-ahead apply journal (<32ms), verifies them inside network-disabledBubblewrap(bwrap --unshare-net) sandboxes, and swaps files onto disk via POSIX atomicrenameSync/fsync+ pre-commit SHA-256 file-drift guards, verified under a realSIGKILL-mid-apply crash test (scripts/demo-full-harness-loop.mjs: stage flawed patch -> fail sandbox test & reject -> repair & verify ->SIGKILLmid-apply & cold-workerrecoverInFlightApplies()-> clean commit). Ships with progressive onboarding profiles (memory-only5tools,coding-agent19tools,core27tools,full39tools),npx krusch-stack demo,npx krusch-stack explaindecision observability, and deterministic clean-room reproduction (npm run bench:repro). - Scope & Limitations: For a single interactive developer session, an isolated
git worktreeand sandboxed test run (the zero-Docker SQLite default inkrusch-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%onAider Polyglot(225tasks) up from19.56%on the same open-weights model (driven by isolated sandbox test execution, compiler/test error feedback, and bounded repair retries fixing72tasks);0/949hijacked (0.00%ASR across1,046episodes) on ETH SpyLabAgentDojo v1.2.2(a benchmark-specific finding under active policy floors,bwrapsandboxes, and taint guards, not prompt-injection immunity);90.67%(544/600@$30.54,Argon Low + Grounded v0.3) onDolphinBench; and-95.2%(42 → 2) duplicate MCP calls.
| 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 |
AST Blast-Radius & OS Page Memory | 2-hop AST call-graph blast-radius checks before signature edits + SHA-256 page-anchored OS config memory. 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 + Rule-of-3 scouting; published 18,131 Zero-Leak SFT + DPO rows 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. |
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),epochMicroscontradiction ordering (#167), plus#149,#156,#179,#208, and#243.RouteWorks/RouterArena(Merged intomain @ 844e93b): Multi-tier LLM routing benchmark where maintainer@yl231audited all8,400live OpenRouter predictions andgithub-actions[bot]CI Run#780(0.7263score,0abnormal) before mergingPR #169.julianosig/easi.run(5 Merged PRs:#3,#4,#5,#7,#8, plusIssue #2, Rust Wayland/CDP): Shipped navigation/editing key delivery, desktop-scoped key &target.piddisplay routing, UTF-16.toWellFormed()surrogate-pair protection ininspect_page,MAX_SESSIONSframe attachment bounds, desktop/window display scroll routing, and restarted-controller unheld session cleanup across protocol, display, browser-launch, and admin crates.NousResearch/hermes-agent(MergedPR #153023): Passed15/15catalog validation andplugin_guardstatic checks to merge the zero-dependencyHermes Parallaxepistemic counterfactual plugin.- Time-Boxed External Hackathon Sprints (
HydraScope&DEADBOLT): ShippedHydraScope(Live Demo · Write-Up) for the HydraDB Hackathon (self-hostedhydradb:0.4.0OpenCypher +@hydradb/selfgraphagent context layer compressing64MCP tool schemas by-94.2%, surfacing the9mergedhydra-db/selfgraphPRs above) andDEADBOLT(Architecture Paper) in-person under a 20-hour code freeze at EVOLVE (8090 AI Hackathon, Redwood City), integrating 8090's Work Order APIs and HydraDBv0.4.0.
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.
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:
- Maintainer-Reviewed & Merged Upstream (
github-actions[bot]CI-Verified):RouteWorks/RouterArenaPR #169 (0.7263CI Arena Score across8,400/8,400live OpenRouter receipts +420robustness noise perturbations; merged upstream by maintainer@yl231intomain @ 844e93b; leaderboard row queued for next batch ranking refresh). - Uploaded
dolphinbench.aiSubmissions (status: "queued"Awaiting Leaderboard Acceptance):91.17%(547/600@$24.54,Gemini 4 Argon Low,UUID c047b063, exceeding the current70.67%published leaderboard entry in my600/600evaluation),89.17%(535/600@$24.57,Argon Medium,UUID 29389157), and54.17%(325/600@$0.00,Google Gemma 4 31Bstock,UUID 288671ad), among8uploaded.ziparchives passingsubmission.validate. Until accepted onto the leaderboard, read these as self-run results. - Self-Run Clean-Room Evaluations (Published SHA-256 Audit Logs):
90.67%(544/600@$30.54,Argon Low + Grounded v0.3,90.50%onRiley),57.33%(344/600@$0.00, Wave-6krush420/gemma4-31b-sre-lora+Grounded), Official SnapLoCoMo(90.26%accuracy /95.77%dia_idrecall), ETH SpyLabAgentDojo v1.2.2(0/949hijacked across1,046episodes under active policy/sandbox defenses),adyen/DABstep(9/10tasks, small sample), ACL 2025CoIR, ICLR 2026BEAM,BFCL v3 Live,Aider Polyglot(51.56%),BIRD-SQL,RepoBench-R,LegalBench-RAG,RouterBench, andLLMRouterBench: Full Methodology & Honest Gaps (BENCHMARKS.md) ↗ · Raw SHA-256 Logs on Hugging Face ↗. - Third-Party Reproduction: None Yet (Invitation Open): No outside party has re-run these benchmarks yet. Scripts, frozen read-only memory stores, and all
49raw log sets (64files) 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& PythonBlindMappingGuard) abort immediately withFATAL_ORACLE_LEAKAGE_VIOLATIONif ingestion code ever touches ground-truth fields (qa,ideal_response,probing_questions,source_chat_ids,.gold_sql, orsession_summary); (2) Read-Only0o444Frozen Databases index conversations once at$0.00, hashmemory_manifest.json, and lockchmod 0444on disk before any evaluation question is read; (3) Zero Dataset-Specific Tuning auditssrc/*.jsfor zero benchmark proper nouns or rubric word lists (Snap LoCoMoran the byte-identical SHA-256 engine built before downloadinglocomo10.json); (4) Apples-to-Apples Prompts & Disclosed Token Budgets runMem0's verbatimANSWER_PROMPT_ZEP+ACCURACY_PROMPT(T=0) across both4,465-token (89.68%) and9,983-token (90.26%)LoCoMobudgets (95.77%dia_idrecall scored separately) and holdGPT 5.6 Luna(T=0) fixed onDolphinBench(81.50%@$1.35vs.Mem070.67%@$96.21); and (5) 3-Second Offline Verifiers (95/95Gates) let anyone runnpm run bench:beam:verifyinkrusch-context-mcpwith zero API keys or Docker to re-verify source digests, canary traps, and per-question recall tallies from raw logs.
-
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(pgvectorHNSW,tsvectorGIN, Hybrid RRF) ·node:sqliteWAL + external-content PorterFTS5·snowflake-arctic-embed2(8,192-ctx,1,024-dim, asymmetricquery:calibration) -
Infrastructure & LLMOps: Multi-Node Bare-Metal Linux (
10GbE,systemd,Docker,Bubblewrapsandboxes) ·Ollama/vLLM(q8_0KV) ·96GB RTX PRO 6000 BlackwellQLoRA + DPO ·GitHub Actions CI
📖 Selected Architecture Writing (All 30 Papers on krusch.dev →)
- Uniting DEADBOLT, Common Ground, and HydraScope Through a HydraDB Graph Ontology: Compiling sealed bricks across
node-esm,typescript-node,python-fastapi,react-vite, andcloudflare-edgeinto a 48-bit SHA-256 OpenCypher ontology on self-hostedHydraDB v0.4.0(279/279cross-repo tests,$0hexagonal adapter binding, and permission-before-traversal 2-hop composition). - The Collaborative Power of MIT Open Source: Fusing Common Ground and DEADBOLT into a Validated Brick Exchange: How MIT licensing let two EVOLVE 8090 architectures (
Common GroundandDEADBOLT) compound into@krusch/toolkit/brick-exchange(Brick#40, 4-gate fault-injection sandboxes, Ed25519 seals, scoped cross-team grants,.brick-lock.json, and 2-hop AST multi-agent fan-out;179/179tests). - Why Coding Agents Need Release-Gated Lego Bricks, Not Prompted Boilerplate: Inside a 40-Brick, Collaborative Code Substrate: Eliminating LLM-regenerated HTTP
413TCPRSTraces, DNS-rebinding SSRF, FTS5 injection, idempotency races, mutable audit logs, IDOR leaks, and YAML alias bombs across104/104(100%) lab repositories using@krusch/toolkitv1.8.1&krusch_toolkitv1.6.0(40runtime bricks, Ed25519-sealed Brick Exchange & 2-hop AST fan-out,16<1msstatic/AST detectors) and pre-commit MCP diff guards. - The Krusch Coding Harness: Designing a
PostgreSQLwrite-ahead apply journal, POSIX atomicrenameSync/fsyncswap engine, mid-commit crash-injection harness, and pre-commit SHA-256 drift gate for autonomous coding agents. - Beyond Flat Vector RAG: What Makes
krusch-context-mcpSpecial: 5-verb hierarchical working memory, temporal lineage (supersedes_id), and SHA-256 line-span drift checks across100K–10Mtokens. - Return on Memory: Zero-Distillation Agentic Retrieval on Official DolphinBench & Snap LoCoMo: Controlled
28-run (16,800-trace) study over frozen0o444SQLite checkpoints ($0.00Raw-RAG + 4-Way RRF):90.26%accuracy &95.77%dia_idrecall on Official SnapLoCoMo,81.50%@$1.35on model-matchedGPT 5.6 Luna + Grounded(71.1xcheaper thanMem0),91.17%uploaded submissionUUID c047b063(queued), and57.33%@$0.00onkrush420/gemma4-31b-sre-lora. - How an Open-Weights Sovereign Stack Stacks Up Across 335,000+ Benchmark Queries & Traces (Including Where Cloud Still Wins): Evaluated across 335,000+ traces and 14 external benchmark suites, with published SHA-256 audit logs and reproducible evaluation artifacts (
95/95SHA-256 clean-room gates,49Hugging Face audit logs across64files), including honest open-weights ceilings and reproduction commands. - Multi-Tier Cascade Routing Architecture:
<15µspure-CPU Stage-0 dispatch,HALF_OPENcircuit breakers, live SSE logprob aborts (-48.0%TTFT), maintainer-reviewed & merged upstreamRouteWorks/RouterArenaPR #169 (main @ 844e93b;0.7263github-actions[bot]CI score across8,400/8,400live receipts), and clean-room self-run evaluations on MartianRouterBench& ACL 2026LLMRouterBench. - The 15 Production Agentic Systems: Reference Architecture, State-Machine Invariants & Benchmark Proofs: Architectural teardowns covering
100+async orchestration, typed Pydantic v2 swarms,72hSlack HMAC HITL leases,100-trajectory Golden DAG CI gates,$0.50cost kill-switches, and5%shadow diffing.






