Joseph Stickel
Applied AI Engineer, Production RAG, Agents, Evals & Private LLM Platforms
jobtrue.ai/u/mrjstickel/applied-ai-engineer
Disable "Headers and footers" in print dialog
Joseph Stickel
Applied AI Engineer, Production RAG, Agents, Evals & Private LLM Platforms
Everett, WA · contact@mrjstickel.com ·
mrjstickel.com · github.com/j-stickel · linkedin.com/in/mrjstickel
Summary
Built Architecture Zero from scratch: a white-label AI platform running four live production instances from one codebase, with hybrid RAG (vector + BM25 + cross-encoder reranking), department-scoped knowledge bases, role-based access, and multi-provider LLM routing with local inference on RTX 5090/3090 hardware. Drove retrieval recall from 40% to a measured 100% through a continuous eval harness, 54 questions, stamped in-app, regression-tested on every change. Every system ships with security gating, cost observability, and fail-closed access boundaries verified by adversarial testing. Bring 13+ years of enterprise support instincts to production AI: when things break at 2am, the monitoring, the postmortems, and the on-call habits are already built in.
Technical Skills
AI / LLM SystemsRAG pipelines (hybrid vector + BM25 fusion), cross-encoder reranking, retrieval eval harnesses, agentic tool use, QLoRA / PEFT fine-tuning, multi-provider routing (Claude / GPT / Ollama) with failover, local inference (7B to 30B-class on RTX 5090/3090), ChromaDB, LanceDB, embeddings, prompt-injection defense
BackendNode.js, Firebase Admin, REST API design, Stripe, SQL, MySQL, SQLite, Alembic migrations
InfrastructureGCP, Vercel, Cloudflare (DNS, tunnels), Docker Compose, Nginx, Let's Encrypt, CI/CD security gating (gitleaks, pip-audit, bandit, trivy), BetterStack monitoring
FrontendNext.js, React, TypeScript, Tailwind, React Native, Expo
Real-TimeFirebase RTDB, Upstash Redis, Server-Sent Events, distributed locking
Experience
Founder & Platform ArchitectJan 2022 - Present
MetaOne Designs (M1D) - Architecture Zero, TheStatic.tv, JobTrue · Remote
- Architected Architecture Zero, a private white-label AI platform running four live production instances from one core codebase - hybrid RAG (vector + custom Okapi BM25 + cross-encoder reranker), department-scoped knowledge bases with role-based access, multi-provider LLM routing (Claude / GPT / local Ollama) with failover, and local inference (7B to 30B-class models on RTX 5090/3090 hardware).
- Drove retrieval recall from 40% to a measured 100% (54-question eval set, stamped in-app) through a continuous evaluation harness: cross-encoder reranking, markdown-section chunking, a full-corpus BM25 leg, and content-addressed delta ingestion that cut re-index costs from ~470 chunks to ~10 per update.
- Built Eco Mode, a federated retrieval mesh: instances query each other's knowledge in parallel with per-caller scoped keys and fail-closed access boundaries - a public bot can never reach private departments; verified by adversarial testing as the least-privileged caller.
- Hardened the fleet with automated CI security gating (gitleaks, pip-audit, bandit, trivy), OWASP Web + LLM Top 10 mapping, and self-audited postmortems - found and closed a real broken-access-control bug, then published the case study.
- Shipped TheStatic.tv solo: a production Web3 streaming platform - 160+ authenticated API endpoints, GCP Livestream RTMP/HLS pipeline, a cron-driven billing engine with Redis distributed locks and viewer-scaled burn rates, and an ERC-721 contract on Ethereum mainnet.
- Shipped JobTrue (jobtrue.ai) solo: an AI job-search SaaS with Stripe billing, LLM fit-scoring with a server-side grounding layer that drops hallucinated resume items before storage, and per-user rate limiting - this resume is served from it.
- Operate everything built: hosting, CI/CD, monitoring, security review, backups, and documentation - single-handedly, in production, for years
Technical Support SpecialistSept 2025 - Feb 2026
Decentraland Foundation · Remote
- Provided technical support for creators and developers building on the Decentraland metaverse platform across Intercom, Discord, and Slack.
- Troubleshot SDK issues, scene deployment problems, wearables, and platform bugs; documented reproductions and coordinated with engineering teams for resolution.
- Leveraged 4+ years of hands-on Decentraland development experience to give expert-level guidance to the creator community.
Technical Support Team Lead / Support Engineer IINov 2022 - Jun 2025
Symetrix, Inc. · Mountlake Terrace, WA
- Hired as Tier 2 support for a professional audio manufacturer; promoted to Team Lead in June 2023.
- Took ownership of NetSuite as primary administrator, rebuilt ticketing workflows and KPI tracking, driving live call answer rate from under 25% to over 80%.
- Redesigned the Return Authorization process mid-2024, cutting turnaround from several weeks to 1-2 weeks.
- Completed a NetSuite ↔ ZOHO integration proof of concept in early 2025.
- Received 2024 Performance Review: Exceeds Expectations on Quality of Work, Takes Initiative, Technical Skills, Dependability, Punctuality, and Attendance.
- Recognized publicly by VP Sales & Marketing for spearheading the call and case routing overhaul.
Senior Support EngineerDec 2018 - Mar 2022
SAP Concur · Bellevue, WA
- Provided Tier 3 enterprise support for SAP Concur expense and travel management products.
- Maintained 98% customer satisfaction across a high-volume enterprise queue.
- Served as Comms on the Incident Response Team for P1/P2 issues, owning customer and stakeholder communication while Tech and Lead drove resolution under SLA pressure.
- Queried client data with SQL to diagnose enterprise expense and travel issues during Tier 3 troubleshooting.
- Mentored and trained team members on technical troubleshooting; built reports and dashboards to track support performance.
Selected Projects
Architecture Zero
mrjstickel.com/projects/azPrivate white-label AI platform | 4 live instances | measured 100% retrieval recall
- Engineered hybrid retrieval: vector search over-fetches a candidate pool, a from-scratch Okapi BM25 leg runs across the full corpus, and a local cross-encoder reranks ~60 candidates to the best 5 - recall driven 40% -> 93% -> 100% by a continuous eval harness (54 questions, stamped in-app).
- Built content-addressed delta ingestion: only new or changed text embeds on updates (~10 chunks instead of ~470), a mid-ingest failure resumes cleanly, and every ingest emits an observability event.
- Designed department-scoped knowledge bases with tiered access and a federated mesh (Eco Mode): parallel peer fan-out, per-caller scoped keys, fail-closed boundaries, graceful degradation when a peer is down.
- Secured the fleet: OWASP Web + LLM Top 10 mapping, CI gates (gitleaks, pip-audit, bandit, trivy), guest budget caps, injection checks, and adversarial self-audits with published postmortems.
Northwind AI (live demo)
northwind.mrjstickel.comWhite-label client demo of Architecture Zero - click it, ask it anything
- Live public deployment proving the white-label model: a fictional company's branded assistant with real department access boundaries - the "View as" switcher shows Sales seeing pricing docs a public visitor is refused.
- Answer quality is eval-driven like everything else: a 35-question set including visitor-voice phrasings, measured at 22/22 passing with zero fabrication - the assistant cites sources and refuses honestly when content is missing.
Kintsugi (private)
Personal AI over a life's knowledge - the private flagship instance
- Mobile-first (React Native/Expo) private assistant over a personal corpus: session history, dated-log chunking with recency weighting, department routing, GPU-tunneled local inference (Cloudflare tunnel to RTX 5090), and the eval discipline the case studies document.
- Private by design - the live public counterpart to walk through is Northwind AI.
JobTrue
jobtrue.aiAI job search platform · Solo build · Stripe billing
- Solved AI resume hallucination structurally: the LLM selects from the user's original bullets by index and application code reassembles verbatim text - a server-side grounding layer drops and logs any fabricated item before storage.
- Identified and remediated a critical SSRF vulnerability in the URL scrape route - blocks private IP ranges, loopback, link-local, and internal hostnames before fetch to prevent cloud credential exposure.
- Shipped Stripe billing (checkout, idempotent webhooks, customer portal), per-user AI rate limiting via Redis sliding windows, fit-score calibration with two-shot anchoring, cost-routed models, and a full admin console with token-cost monitoring.
- Built the application pipeline tracker: frozen resume + cover letter snapshots per submission, interview rounds, contacts, and fit-score history.
MySQL Client & SQL Trainer (private)
Full-stack database tool · Solo build · Mobile + API
- Built a MySQL management client: connect to live databases, browse schema and indexes, and run and save queries from a mobile app.
- Integrated an AI assistant (Claude) that reads your actual schema and writes or explains SQL against it on request - constrained to read-only queries.
- Engineered query safety: Fernet-encrypted credentials at rest, read-only enforcement with an explicit write-mode opt-in, and multi-statement blocking on every connection.
- Added a 40-challenge SQL trainer on a separate in-memory SQLite database, grading answers via automated row-set comparison - practice isolated from production data.
TheStatic.tv
thestatic.tvProduction Streaming Platform · Solo build · 160+ API endpoints
- Engineered GCP Livestream API integration with RTMP ingest, tier-based HLS transcoding (720p -1080p60), and CDN delivery via Cloudflare.
- Architected distributed Spark Economy with Redis distributed locks (120s TTL) and viewer-scaled burn rates (1x → 50x).
- Implemented Polygon crypto payments with verification windows and a 6-tier subscription system.
- Built multi-auth system: MetaMask (EIP-191), Decentraland signed fetch, Firebase custom tokens.
- Deployed 13 production cron jobs, batched VOD archival, and admin audit logging.
What I Deliver
· Production AI systems end-to-end: RAG pipelines with measured retrieval quality (40% -> 100% recall, eval-stamped), agents with tool use, guardrails, and cost observability
· Private / self-hosted AI platforms - your data, your brand, your infrastructure (Architecture Zero: four live instances from one core)
· Retrieval engineering: hybrid vector + BM25 fusion, cross-encoder reranking, section-aware chunking, delta ingestion - fixed at the corpus, not patched at the prompt
· Continuous AI evaluation: question sets, regression floors, measured before/after on every change
· Multi-provider LLM architecture: Claude, GPT, and local Ollama with failover; local inference on 24GB-class GPUs
· Federated / multi-tenant AI: department-scoped knowledge, role-based access, fail-closed boundaries - adversarially self-tested
· AI security in production: OWASP LLM Top 10 mapping, prompt-injection defense, CI security gates, published self-audit postmortems
· LLM fine-tuning through the full QLoRA lifecycle - and the senior call of when NOT to fine-tune
· Full-stack SaaS shipped solo: auth, Stripe billing, admin tooling, transactional email
· P1/P2 incident communication and SLA-driven operations - 13+ years of enterprise support instincts wired into everything above
Education
BS, Information Technology Management (BSITM)Jun 2026 - Expected 2027
Western Governors University (WGU)
Returning student completing the degree while building and operating production AI systems full-time.
Built with JobTrue · Create your free resume →