Erik Stephens

Erik Stephens

SW Engineer - Backend/Systems

Cultivating practical software systems with equal parts architecture, experimentation, and operational care. I enjoy building things that are useful, resilient, and quietly sustainable: from backend platforms & observability pipelines to search systems & AI-assisted workflows.

Here is a prompt to help evaluate me based on publically visible projects. Feed it to your LLM of choice with any other relevant context to see how well I fit.

Featured

All
FORGEJO MEMORY accepted closed OPTIONAL supervisor telemetry mgmt ui rejected findings findings issue PM DEV QA human forgejo wiki AGENTS.md project wiki

FJX

Agent Orchestration • Deno, MCP, Forgejo

An agent-accelerated forge where humans and LLM agents collaborate on software using the same artifacts.

SOURCES SINKS journald file parse filter aggregate enrich tap debug tap sample OpenSearch prod file analysis

LogBus

Event Processing • Rust (ported from TypeScript)

A programmable event-pipeline engine — YAML for the shape, JavaScript and WASM for the weird cases, backpressure all the way to the source.

LAPTOP MOBILE USB DRIVE SERVER BUCKET S3 · R2 · B2 ›› CONTENT-ADDRESSED files keyed by hash — automatic deduplication TOMBSTONED DELETES removals propagate safely, never silently resurrect SAFE CONFLICTS two edits? both versions survive — no overwrite REMOVABLE DEVICES 1st-class USB & external drives, mounted or not TRANSPARENT LAYOUT files map 1:1, readable without DropBear

DropBear

Storage Utility • Go, SQLite, S3

A self-hosted file-sync daemon that pairs local folders with any S3-compatible bucket — with a daemon, web dashboard, and menu-bar app.

RETRIEVER PROTOCOL SWAPPABLE BACKENDS elasticsearch postgres query BM25 · inverted index dense kNN · HNSW RRF fusion OPTIONAL cross-encoder rerank eval harness ir_measures nDCG@10 · R@100 · MRR BEIR qrels

SearchLab

Search • Python, Elasticsearch, pgvector

An offline relevance lab for hybrid search — BM25, dense vectors, RRF fusion, and cross-encoder reranking, measured against public BEIR benchmarks instead of vibes.