Sheet 02 · Assemblies & details

Projects

Two things I am building now, in detail — and the professional work that came before them.

Part 01 · Founder · 2025 — now · Live, 30+ users nextloom.ai ↗

nextloom.ai

An AI-powered job search assistant for software engineers. It finds relevant roles, tailors the resume and cover letter for each one, helps prepare for interviews, and applies through each company’s own hiring system. I designed the product, architected the system, built it and run it — alone.

Clients

Browser extension

CLI

Core platform

Hono API Effect.ts · DDD · CQRS

LLM workflows LangGraph · LangChain

State PostgreSQL checkpoints

Hiring systems

Lever

Greenhouse

Ashby

Workday via Jobo

Fig. 1 — nextloom.ai architecture

Problem: Applying well is slow: every role wants a tailored resume, a cover letter and a different application form.
Approach: Stateful LLM workflows draft the materials; native ATS APIs submit them first, with a paid layer only where no API exists. A person approves what goes out.
Key decisions: LangGraph with PostgreSQL checkpoints so long workflows resume safely · DDD and CQRS-style separation to keep the domain clear · Effect.ts for typed errors across a pnpm monorepo.
Stack: TypeScript, Hono, Effect.ts, LangGraph, LangChain, PostgreSQL, Vercel
Status: Live, 30+ users

Part 02 · Open source · v0.9.0 · 2026 GitHub ↗Docs ↗

Lingtai

A self-hosted tool that runs a repeatable coding-agent loop on your GitHub issues. It picks up labeled issues one at a time, has an agent work each one in an isolated worktree, runs your repository’s checks, and lands work that passes. When a run needs a person, it stops and records why.

Input

GitHub issues labeled

Control

Recipe labels · agent · checks · approvals

Queue & board

Run

Isolated worktree

Coding agent guarded run

Verify

Repo checks

Outcome

Land

Wait for review reason recorded

Fig. 2 — Lingtai run loop

Problem: Coding agents are useful, but running them unattended on a real repository needs structure, isolation and a clear stop condition.
How it works: lingtai init → ticket new → doctor → start. A local board shows queued, running, waiting and landed work.
Guardrails: One worktree per run, a filtered environment and a recording hook; optional human approval before merge; honest list of known limitations.
Stack: CLI for macOS and Linux, SQLite by default or Postgres, GitHub App
Status: v0.9.0, first public release (Aug 2026)

Professional work

Full resume →

Part 03Oracle

Helium

Docker containerization platform and CLI, built solo from idea to adoption. Multiple product teams, 800+ developers; setup from 2+ hours to 15 minutes.

Part 04Oracle

LLM-as-judge

Evaluation platform for chatbot response quality, built solo. Phase 1 in production; significantly increased user trust in LLM results.

Part 05Fuhu

Event-sourced user management

Cassandra event store with replay audit trail for the Nabi platform — 1M devices, 2M child profiles. Led a team of 6.

Part 06Fidelity International

First iOS investment app (UK)

Led development of Fidelity’s first iOS investment application for the UK market.

Part 07Neusoft · China Unicom

Partner relationship management

PRM platform and service provider management system. Led a team of 10.