Building tools for the era when AI writes the code.
工欲善其事，必先利其器。
agents · evals · macOS menu bar apps · digital twins | e/acc · 加速 | @seasonsolt
Architect for 10+ years — fintech and e-commerce backends in Java, payment chains that held 10K+ QPS on peak days. Now a technical partner at an AI startup, leading a 10-person team building agent platforms: runtime, MCP hub, skills hub, model gateway. These days I mostly write Python and Swift — agents that code, evals that judge them, and small native apps that keep the whole loop visible from the macOS menu bar.
Not "move fast and break things" — move fast and measure everything.
A few things I believe:
- Eval-driven development. Write the eval before the prompt — benchmark, golden set, LLM-as-judge. The eval is the spec.
- Ambient tooling beats dashboards. Agent status belongs in the menu bar, not in another browser tab.
- Acceleration is a practice, not a vibe. Ship the loop, tighten the loop, repeat.
Same discipline the whole way: systems that hold under load. The load is just measured in tokens now.
| project | stack | what it is |
|---|---|---|
| twin | Python |
Personal digital twin that answers, speaks and appears like the person — every claim cited from their own words |
| tachi | Swift |
Native macOS menu bar companion for live AI coding sessions, Claude usage & Codex quota |
| lancer | Java |
Debug one JVM microservice on your laptop against a shared Kubernetes cluster — only your requests hit your breakpoint |
| skills | Python |
Verifiable, boundary-aware agent skills for creation, knowledge work and software engineering |
| edd-skill | Python |
A skill for eval-first AI coding workflows |
| sac-agent | Python |
Teaching-oriented terminal software-engineer agent — Textual, LangChain, DeepAgent |
Also: sac-agent4j — the same agent ideas in Java, for my backend roots.
The names: sac is Stand Alone Complex, tachi a Tachikoma — the think-tanks that synced memories nightly. Agent tooling had a spec in 2002.
📦 More cargo in the hold
- agent-lab — FastAPI coding agent with a local Langfuse stack, where the evals get run
- cqrses-order — CQRS / Event Sourcing order system demo (the Java roots run deep)
- ddia-skill — Designing Data-Intensive Applications as a study skill
- e-acc.ai — you already know what this one is about (live)
- ritual-screen — sacred token-consumption visualization for vibe coding
- Built a vertical-agent platform — agent runtime, MCP hub, skills hub, model gateway — that took AI product delivery from idea to production 2–5× faster.
- Shipped a legal-AI document workflow to overseas law firms and funds: 90%+ extraction accuracy on 200+ page filings at P99 ≈ 5s/page.
- A decade of load-bearing backend work: order/payment/clearing systems, 10K+ QPS promotion peaks, cross-border payment fees cut 0.5% → 0.2%.
DDIA, chapter by chapter — ddia-skill is me rereading it with an agent · eval philosophy, including where "the eval is the spec" breaks down · writing about all of this in 中文 — I think in two languages and debug in one · the avatar — start with Salinger
Worth an email: agent platforms in production, eval harness design, or a menu bar app that should exist — seasonsolt@gmail.com
「唯快不破」 — nothing beats speed.