~/aspesi.dev

Projects — four things worth talking about.

Ordered by what I'd actually want to talk about, not by date. A shipped iOS venue-discovery app I co-founded and lead, a conversational lead-qualification agent I built for a car-buying business, the multi-agent review harness I run over my own diffs, and open-source contributions to the workspace orchestrator I develop on. One has a full write-up; the rest link out where the source is public.

4 projects · ranked by significance

  1. 2024 — now
    ~/projects/on-the-grid

    On The Grid

    shipped · App Store

    A native iOS app for discovering and saving places worth going — save a spot from Instagram, TikTok, or Google Maps in one tap, then meet people nearby who saved the same places.

    Every place worth going is already sitting in a screenshot, a TikTok, or a friend's link — just nowhere you can act on it. On The Grid pulls the places you save from Instagram, TikTok, Google Maps, and Beli onto one map of your own, then turns them social: you find people nearby who saved the same spots and make plans together. Deliberately a discovery-and-plans app, not a dating app.

    ios · swift

    read the case study → repo ↗

  2. 2025 — 2026
    ~/projects/otto

    Otto

    shipped · client engagement

    A LangGraph.js conversational agent for a car-buying service that walks a shopper from their first message to a qualified sales lead — pulling live dealer pricing and handing the lead to a CRM.

    Car buyers show up knowing exactly what they want, a rough idea, or nothing at all. Otto handles all three as a conversation instead of a web form: it extracts what it can from each message — location, vehicle, terms, contact — routes on the buyer's intent, and only asks for what's still missing. Once it has enough to identify a real purchasable deal it fetches live lease, finance, and cash pricing from dealer inventory and submits the completed lead to the client's CRM. The graph is an extract → route → act design over conversation state, with a Zod schema driving single-pass extraction and a LangSmith eval suite scoring extraction, intent, and stage progression.

    typescript · langgraph.js · openai · zod · postgres · redis · langsmith

    repo ↗

  3. 2026 — now
    ~/projects/paseo

    Paseo

    active · contributor

    Open-source contributor to Paseo (~11k★), a self-hosted orchestrator that runs many AI coding agents in parallel from one interface — a merged iOS fix plus PRs for session-forking and workspace-pinning.

    Paseo runs a local daemon that manages Claude Code, Codex, Copilot and other coding agents as parallel processes, with iOS, desktop, web, and CLI clients connecting to it — it's the workspace layer this very site is developed on. I contribute against its TypeScript daemon and Expo/Electron clients: a merged fix that made markdown links tappable on iOS (#1334), plus proposed PRs for forking a session into a new tab (#2283) and pinning workspaces to the top of the list (#2104). Author and maintainer is Mohamed Boudra; I'm a contributor, not the author.

    typescript · node · expo · electron · websocket

    repo ↗

  4. 2025 — now
    ~/projects/claude-harness

    Claude Code harness

    active

    A personal multi-agent review and test-coverage command suite — two independent reviewers, then a judge that verifies their findings against the real source.

    A single reviewing agent is confidently wrong often enough to train you to ignore it. So two reviewers — Claude Code and Codex CLI — read the same diff with no knowledge of each other, and a judge subagent re-opens the files to verify every finding against real code, dropping anything it can't reproduce before a human sees it. It reads a per-repo REVIEW.md for local rules and auto-detects project type to adapt its prompts. False-positive rate mattered more than recall here: the suite is tuned to say less and be right, because a review nobody trusts is worse than no review.

    claude code · codex cli · subagents · bash · markdown rule files

    repo ↗