hi, i'm rein.
I'm a software engineer with 8+ years building and shipping production web apps — React, Vue, Next.js, TypeScript — to enterprise customers like FedEx, Amazon, and CoreWeave, with around $700k in annual client cost savings along the way.
These days I'm focused on applied AI engineering: building production-grade systems on LLM APIs — a provider-agnostic agent runtime, multi-agent diagnostics on synthetic data, and evaluated RAG pipelines.
I'm looking for a role I can contribute to from day one on what I know best — full-stack, front-end-leaning product engineering — while growing into the applied AI and agent work I've been building toward.
seattle, wa · open to front-end / ai engineering roles
coreweave
A production-grade AI analyst for ecommerce workspaces: three specialist agents — monitor → diagnose → recommend — over MCP, streaming full reasoning provenance (every tool call, metric delta, and hypothesis) to the UI via SSE. Hardened for production — OAuth 2.1 / PKCE auth, per-call timeouts with retry ladders and budget ceilings, CI/CD, and an eval harness (221 tests, rubric-as-judge over golden cases) so agent quality is measurable.
A TypeScript monorepo of composable agent packages: a runtime with bounded tool-use loops and structured-output-with-retry, model adapters (Claude, OpenAI, local Gemma), plus RAG with swappable embeddings/vector stores and episodic conversation memory. Published to npm and reused across my agent projects, with deterministic replay evals (precision@k / recall@k) for regression-testing trajectories across model and prompt changes.
A self-hosted RAG agent with an interactive chat CLI: indexes notes into pgvector (HNSW) on Supabase Postgres and answers grounded questions via a local Gemma model, with persistent cross-session conversation memory and profile-based grounding. Built around eval discipline — precision@k / recall@k scorers and full per-step trajectory capture to drive data-driven fine-tuning.
A routing engine that optimizes for flat routes over fast ones — a hand-rolled A* over an OpenStreetMap + elevation graph, with admissibility and edge cases proven by 130 tests.