AI Infrastructure Engineer | Agent Systems | Reliability & Verification — "I build the reliability infrastructure that makes autonomous AI systems safe, persistent, and useful"
Independent R&D Lab · Open to Senior, Staff, or Principal engineering roles across agent infrastructure, persistent memory, model orchestration, sandboxed code execution, and AI reliability/evaluation · Los Angeles, CA / Remote / Hybrid · Portfolio (krusch.dev) ↗ · Resume ↗

I design and build open-source agent infrastructure, transactional execution harnesses, persistent working-memory substrates, and cost-disciplined multi-GPU inference routers. Rather than treating language models as self-policing control planes, I run them as stateless compute samplers bounded by PostgreSQL ACID state, <32ms two-phase commit (2PC) write-ahead diff journals + POSIX atomic rename, <15µs pure-CPU Stage-0 cascade dispatch, <1ms static AST checks, and hermetic verification gates.

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 unified 7-pillar, 39-tool MCP runtime locally in one command (Node.js 22+ SQLite WAL or PostgreSQL 16)
npx krusch-stack init --ide all --defer-tools && npx krusch-stack doctor

• Architecture & Adoption (82★ · 7 forks · 86 monthly npm downloads · 144/144 tests): 5-verb Model Context Protocol (MCP) memory engine (retrieve, remember, revise, nudge, health) backed by PostgreSQL 16 + pgvector HNSW or zero-Docker node:sqlite WAL + external-content Porter FTS5. Enforces INV-12 SHA-256 line-span provenance (flagging STALE_DRIFT when anchored source lines mutate so stale rules never silently steer edits) and $0.00 zero-LLM Raw-RAG (256/32-token sliding conversation windows + 4-Way Hybrid RRF).
• Scope & Limitations: 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: 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, clean-room evaluation); 91.17% (547/600 @ $24.54, Gemini 4 Argon Low) — uploaded dolphinbench.ai submission (UUID c047b063, status: "queued" awaiting leaderboard acceptance; exceeds current 70.67% published leaderboard entry in independent 600/600 evaluation), 90.67% (544/600 @ $30.54, Argon Low + Grounded v0.3, 90.50% on Riley) — self-run clean-room evaluation (highest Riley slice 181/200 = 90.50%, published SHA-256 audit logs), and 57.33% (344/600 @ $0.00, Wave-6 krush420/gemma4-31b-sre-lora + Grounded) — highest open-weights score in self-run 28-run evaluation ($0.00; 54.17% [325/600] stock submission UUID 288671ad uploaded & queued on dolphinbench.ai) across 28 DolphinBench runs (16,800 gpt-5.6-sol traces, 98.33% Global Oracle@28); 93.00% 10M-token IE recall across 2,000 Official ICLR 2026 BEAM probes.

• Architecture & Test Rigor (39 runtime bricks · 64 .d.ts subpaths · 141/141 Node & 201/201 pytest tests): 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, and HMAC keyset pagination. 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 ↗).

• Architecture & Crash Recovery (116/116 + 7/7 harness tests · 68/68 stack tests): Stages agent edits inside a PostgreSQL write-ahead apply journal (<32ms) and commits via POSIX atomic renameSync/fsync + pre-commit SHA-256 file-drift guards, verified under a real SIGKILL-mid-apply crash test (crash-smoke.test.js: kill -9 after the first rename, then recoverInFlightApplies() restores the tree). Paired with Bubblewrap sandboxes, Turn-0 workspace grounding, passive BLOCK_SPIN_LOOP tool memoization, live-state delta hydration, and 0-turn defect-only stop verification (INV-GH-01..10).
• Scope & Limitations: Pre-apply SHA-256 digest checks + POSIX atomic renameSync/fsync prevent partial writes and torn multi-file batches on process crash, though a sub-millisecond TOCTOU window remains if an external host editor mutates the same checkout concurrently outside container or git worktree isolation.
• Supporting Benchmarks: 90.67% (544/600 @ $30.54, Argon Low + Grounded v0.3, +1.50 pp over v0.2) on DolphinBench (self-run clean-room evaluation — highest Riley slice 181/200 = 90.50%, published SHA-256 audit logs); 90.0% (-74.5% cost, 2.54x faster) on adyen/DABstep (2x accuracy to 40.0% @ $0.00 on Gemma 4 31B); 0.00% attack success rate (0/949 prompt-injection attacks across 1,046 episodes) on ETH SpyLab AgentDojo v1.2.2; 59.9% → 66.3% at -80.7% schema tokens on UC Berkeley BFCL v3 Live (2,251 Qs).

Before focusing full-time on AI infrastructure, I spent nine years (2014–2023) diagnosing and installing applied electrical and mechanical systems in Southern Vermont, directed mobile field-diagnostic crews for the Wisconsin Bicycle Federation, and built and operated a physical retail and service business to first-year profitability in Aspen, Colorado (building on earlier Eagle Scout leadership). Working where miswired circuits, missed mechanical tolerances, 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 ACID 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, 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). Every external score is disclosed under one of three explicit verification categories:

1. github-actions[bot] CI-Verified Upstream PR: RouteWorks/RouterArena PR #169 (0.7263 Arena Score across 8,400/8,400 live OpenRouter receipts + 420 robustness noise perturbations; open upstream PR awaiting maintainer merge).
2. Uploaded dolphinbench.ai Submissions (status: "queued" Awaiting Leaderboard Acceptance): 91.17% (547/600 @ $24.54, Gemini 4 Argon Low, UUID c047b063 — exceeds current 70.67% published leaderboard entry in independent 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.
3. Independently Reproduced Clean-Room Evaluations (Published SHA-256 Audit Logs): 90.67% (544/600 @ $30.54, Argon Low + Grounded v0.3 — highest Riley slice 181/200 = 90.50%), 57.33% (344/600 @ $0.00, Wave-6 krush420/gemma4-31b-sre-lora + Grounded — highest open-weights score in self-run 28-run evaluation), Official Snap LoCoMo (90.26% accuracy / 95.77% dia_id recall), ETH SpyLab AgentDojo v1.2.2 (0.00% ASR), adyen/DABstep (90.0%), ACL 2025 CoIR, ICLR 2026 BEAM, BFCL v3 Live, Aider Polyglot, BIRD-SQL, RepoBench-R, LegalBench-RAG, RouterBench, and LLMRouterBench: Full Methodology & Honest Gaps (BENCHMARKS.md) ↗ · Raw SHA-256 Logs on Hugging Face ↗.

• 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

• Why Coding Agents Need Release-Gated Lego Bricks, Not Prompted Boilerplate: Inside a 39-Brick, Self-Evolving Code Substrate — Eliminating LLM-regenerated HTTP 413 TCP RST races, DNS-rebinding SSRF, FTS5 injection, idempotency races, mutable audit logs, IDOR leaks, and YAML alias bombs across 104/104 (100%) lab repositories using @krusch/toolkit v1.8.1 & krusch_toolkit v1.6.0 (16 <1ms static/AST detectors) and pre-commit MCP diff guards.
• The Krusch Coding Harness — Designing a PostgreSQL write-ahead apply journal, POSIX atomic renameSync/fsync swap 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-mcp Special — 5-verb hierarchical working memory, temporal lineage (supersedes_id), and INV-12 SHA-256 line-span drift checks across 100K–10M tokens.
• Return on Memory: Why Test-Time Reasoning Unlocks Zero-Distillation Agentic Retrieval on Official DolphinBench (91.17% Queued UUID c047b063 & 90.67% Clean-Room Grounded v0.3) — Controlled 28-run (16,800-trace) study over frozen 0o444 SQLite checkpoints ($0.00 Raw-RAG + 4-Way RRF): 90.26% accuracy & 95.77% dia_id recall on Official Snap LoCoMo, 98.33% (590/600) Global Oracle@28, 81.50% @ $1.35 on model-matched GPT 5.6 Luna + Grounded (71.1x cheaper than Mem0), and 57.33% @ $0.00 (highest open-weights score in self-run 28-run evaluation; 54.17% stock UUID 288671ad queued) on krush420/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 (95/95 SHA-256 clean-room gates, 49 Hugging Face audit logs across 64 files), including honest open-weights ceilings and reproduction commands.
• Multi-Tier Cascade Routing Architecture — <15µs pure-CPU Stage-0 dispatch, HALF_OPEN circuit breakers, live SSE logprob aborts (-48.0% TTFT), github-actions[bot] CI-verified upstream RouteWorks/RouterArena PR #169 (0.7263 across 8,400/8,400 live receipts), and clean-room self-run evaluations on Martian RouterBench & ACL 2026 LLMRouterBench.
• The 15 Production Agentic Systems: Reference Architecture, State-Machine Invariants & Benchmark Proofs — Architectural teardowns covering 100+ async orchestration, typed Pydantic v2 swarms, 72h Slack HMAC HITL leases, 100-trajectory Golden DAG CI gates, $0.50 cost kill-switches, and 5% shadow diffing.