platform architecture & systems engineering

Payments, auth, and platform infrastructure — built to hold.

I design and build backend systems where multiple moving parts must work together reliably. Rust, Go, Python. Small teams, hard problems, clear business impact.

What I work with

Rust
Systems, FFI, async engines, cross-platform
Go
Service architecture, Lambda platforms, CLIs
Python
Tooling, automation, data pipelines
AWS
Lambda, Cognito, RDS, SQS, EventBridge, EKS
Databases
PostgreSQL, SQLite, Aurora, schema design
Infrastructure
Terraform, Docker, Kubernetes, LocalStack
Auth & Payments
OAuth2/OIDC, Cognito, multi-provider payments
CI/CD & Tooling
GitHub Actions, Ansible, Packer, custom CLIs
Protocols
JMAP, HTTP, SSE, RFC implementations

What I've done

Featured work

Neverlight Mail

One engine, four clients. Native JMAP mail across platforms.

A from-scratch email engine implementing RFC 8620 and RFC 8621 (JMAP Core and JMAP Mail) directly — no IMAP, no SMTP, no protocol abstraction layers. The engine handles connection management, delta sync, full-text search, and offline caching through a single Rust crate.

Four independent frontends consume the same core: a COSMIC desktop client for Linux, a ratatui terminal client for Linux and macOS, and native SwiftUI clients for macOS and iOS. The Swift clients bridge to the Rust engine via swift-bridge FFI — real native on both sides, no Electron, no webview.

Architecture
neverlight-mail-core (Rust)
  └ JMAP client — RFC 8620/8621, reqwest, SSE push
  └ Cache layer — SQLite, FTS5, async handle pattern
  └ Body engine — mail-parser, HTML→markdown

neverlight-mail (COSMIC/iced) — Linux desktop
neverlight-mail-tui (ratatui) — Terminal, Linux + macOS
neverlight-mail-macos (SwiftUI + swift-bridge FFI) — macOS + iOS
Rust Swift JMAP SQLite FFI ratatui SwiftUI COSMIC

PGTOP

Postgres performance monitoring in your terminal.

A Ratatui-based terminal interface for PostgreSQL performance monitoring. Built with Cornucopia-generated typed queries, it surfaces slow queries, locks, active transactions, and connection state with minimal overhead. The kind of tool you leave running in a tmux pane when you're debugging why prod is slow.

Rust PostgreSQL ratatui Cornucopia

GAP — Game Agent Protocol

Multi-agent LLM system that plays Diablo cooperatively with humans.

A real-time AI companion for Diablo 1 (via DevilutionX) powered by local LLM inference. 16 specialist agents — combat, healing, loot, shopping, exploration, chat — evaluate game state through a compact DSL protocol and vote on actions via weighted council. Constraint satisfaction solvers validate positioning and resolve conflicts. Grammar-constrained output (GBNF) guarantees valid commands — no hallucinated actions.

Persistent SQLite memory gives each companion character its own personality, episodic memories, and learned strategies across sessions. A Rogue remembers kiting tactics; a Warrior remembers face-tanking failures. ~9,000 lines of Python orchestration, 682-line DSL parser, custom C++ game engine hooks.

Architecture
DevilutionX (C++) — game state via Unix socket
  └ DSL protocol — 100-200 bytes/tick, 80-120 tokens

Orchestrator (Python) — 16-agent council + CSP solvers
  └ Combat, Healing, Loot, Spell, Shopping, Chat, ...
  └ Spatial CSP, Equipment CSP, Coordination CSP
  └ SQLite memory + per-character personality

Ollama (local GPU) — qwen2.5:3b / llama3.1:8b
  └ GBNF grammar → guaranteed valid DSL output
Python C++ LLM Multi-Agent GBNF CSP SQLite Ollama DSL

Get in touch

jason@neverlight.com

GitHub