The engineer at the intersection AI deployment is short on.
I'm a one-person AI software foundry who ships production agentic / MCP / RAG systems — and I bring 15 years deploying enterprise software into hard, regulated reality (Toyota, MAN Energy Solutions, Capgemini). That second half is the part most AI engineers don't have, and it's exactly what a forward-deployed role screens for. Every claim below is live and publicly inspectable.
Ships production AI
Open-source MCP server on npm, a multi-provider agent platform with an LLM-as-judge eval gate, RAG pipelines, self-healing CI — published, not slideware.
15 years deploying into reality
Embedded with business stakeholders at Toyota, MAN, Capgemini — frustrated-stakeholder requirements and messy data turned into delivered, supported systems.
What I've shipped
Open source unless noted. Links are live — inspect the commits, the tests, and the design decisions yourself.
abap-mcp — MCP server for SAP ABAP
npm · open source + eval harnessGives AI coding agents offline ABAP analysis, Clean-Core readiness checks, and RAP scaffolding — no SAP system needed. I designed a dual-parse Cloud-readiness diff that separates real migration blockers from pre-existing defects, and a self-validating RAP scaffolder (generator and linter share one parser; every output round-trips through abaplint at Cloud level). 7 tools, 80 tests, a weekly auto-upgrade CI — and a public LLM-as-judge eval harness (deterministic golden oracle + agent-under-test + judge, scored on grade accuracy, category precision/recall, Cohen's κ, and a failure-mode taxonomy). That harness is the exact artifact forward-deployed loops ask for.
Multi-vendor AI agent platform
agenticAn autonomous triage → plan → execute → maker-checker review → auto-merge loop with fallback ladders across 5 LLM providers, gated by a prompt-regression eval suite: deterministic offline stub, LLM-as-judge scoring, a 16-code failure-mode taxonomy, and a token-budget circuit breaker. Changes can't merge without passing the gate.
Property due-diligence app
liveOne address → a web-grounded, cited, weighted-risk report across 36 dimensions (SSR on Vercel, Gemini + Search grounding, Zod-validated structured output). I engineered a self-healing pipeline: runtime errors → Postgres (deduped) → daily AI-triage cron → gated, source-only auto-fix PRs behind a kill-switch.
RAP Dojo — Clean Core academy
liveA 17-module interactive ABAP→RAP academy with a live in-browser linting loop and an always-mounted AI coaching rail. Built on a provider-agnostic AI router (multi-model, rate-limited, token-budget circuit breaker) grounded by a self-hosted SAP-docs MCP.
mcp-kit — MCP toolkit & tool-lint
open sourceA hardened MCP starter (TypeScript + Python/FastMCP) with dual transports, bearer auth, a retryable-error taxonomy, and a novel tool-description lint that treats model-facing docs as lintable code (hard-fails CI on credentials-in-inputs).
Tether — phone-native AI terminal
native Kotlin
A native Android AI-coding terminal (vendored Termux engine + a cross-compiled SSH client)
that runs the whole fleet from a phone. When the NDK couldn't host on ARM64 and box64 segfaulted its
clang, I cross-compiled a prebuilt libtermux.so with clang+lld — low-level work no agent
hands you for free.
How I work
I'm an AI-augmented engineer — I orchestrate agent fleets to ship at high leverage, and I can read, extend, and defend every line they produce. Happy to demonstrate live, no AI assist.
Reliability & eval discipline
LLM-as-judge eval gates, failure-mode taxonomies, token-budget circuit breakers, prompt-injection hardening, self-healing error→triage→auto-PR pipelines.
Full-stack, in their stack
TypeScript / Next.js 16 / React / Vercel (AI SDK + AI Gateway) + Python + ABAP. I solo-operate ~50 repos and ~13 live apps on cost-guardrailed, mostly-free-tier infra.
Deploy into messy enterprises
15 years embedded with business stakeholders — discovery with the frustrated CIO, mapping legacy + compliance constraints, a thin guardrailed first slice, a measurable win, then expand.
Enterprise delivery — at scale, customer-facing
Architected A2A interfacing on SAP PI/PO 7.5 (EDI/IDocs) for scheduling, orders, shipping and invoices; ran daily production support alongside the business. Delivered a sales-process Kaizen saving ~C$100k/year and automated the annual pricing-review workflow.
Built Fiori approval apps (PO / RFQ / PR) with custom multi-step workflows; lifted the annual OTACE client-satisfaction score from 3.7 → 4.1 / 5.
Migrated and optimized 2,500+ custom developments during an EHP8 / Suite-on-HANA upgrade; rollouts across Croatia, Brazil and Thailand.
Implemented master-data governance workflows in a multi-client landscape (ALE/IDoc/EDI); harmonized and de-duplicated master data across 75,000+ customers and 225,000+ suppliers.
Delivered an SAP R/3 implementation plus post-go-live support; BAPI goods-movement interfaces, LSMW/BDC data conversions.
Let's talk forward deployment.
I'm open to Forward-Deployed Engineer, Solutions Engineer, and Applied-AI roles — based in Ontario, Canada and open to US / UK / EU with sponsorship (TN-eligible).