Recently Shipped
Production highlights from 5 projects I've been building — construction SaaS, training-data pipeline, agent messaging, email automations, and a mobile Hermes console. Sourced from each repo's SESSION_SUMMARY.md and README.md.
PocketForeman
Mobile-first construction project management for general contractors.
AI engineer role + training workflows
New ai_engineer role runs Pi coding harness workloads in Vercel Docker sandboxes — owner briefs become training-data workflows with live run status, logs, and artifacts.
Magic link signup + landing funnel
Direct signup via Resend magic links, marketing landing page with UTM tracking, PostHog analytics, and beta access request pipeline.
Budget signoffs with audit trail
Owners add line items; clients approve each one. Estimate/Finalized badges, auto-recomputed totals, and an immutable audit log for legal traceability.
Beta ops layer
Slack alerts for signups and project events, invite CLI, photo/file uploads via Vercel Blob, and role-scoped data for owners vs clients.
Field Pipeline
Open-source training-data pipeline — field expert briefs in, export-ready artifacts out.
Field-sourced, not engineer-sourced
Domain experts supply raw briefs and label rules; AI engineers write Pi prompts. The pipeline structures and generates datasets from people who actually do the work.
Vercel Sandbox + Pi + your LLM
Isolated microVM runs Docker with Pi in print mode. Smoke tests finish in ~1–2 minutes. Bring your own Anthropic or OpenAI key — no model vendor lock-in.
YAML workflows + CLI
fieldInputs and engineerPrompt in workflow YAML; validate and run via CLI. Example starters for construction, healthcare, and retail.
Portable export layer
Manifest JSON, Hugging Face dataset-card stubs, and OpenAI fine-tune envelope helpers — JSONL, CSV, or custom formats declared in workflow output spec.
Extracted from PocketForeman
Open-source core of the training-data engine prototyped in PocketForeman. MIT-licensed engine any industry can embed without the mobile app.
AMP
Agent messaging platform — tag @claude, @codex, or @opencode in channels.
Team members + org invites
Owner/admin invite flow with role-based permissions. Fixed multi-tenant demo-org leak — LastMyle and PocketForeman now run as separate orgs with verified invite E2E.
CLI harness for all three agents
@claude, @codex, and @opencode run real CLI binaries in Vercel sandboxes with CLI badges, Stop button, and harness persistence across logout.
BYOK + file uploads
Org inference credentials in Settings (encrypted at rest). PDF and image attachments work in both Eve and CLI harness paths via private Vercel Blob.
Durable threads + org memory
Cross-session thread history, org-scoped memory tools (remember/list/forget), and Eve sync to eliminate ghost sessions after canceled workflows.
Agent discovery (Phase 0)
Server-rendered homepage, robots.txt, sitemap, /llms.txt, and JSON-LD for agent crawlers.
EB Automate
Email-first AI automations for small businesses — $9/mo.
Multi-user workspace (Phase 0b)
Owner invites members; shared automation threads with per-user attribution. RBAC: members post and automate; owners approve email drafts and manage billing.
Learning memory (Phases 0–2)
Team USER/MEMORY profiles, Postgres FTS search across past automations, prefetch on each turn, and async request summarization.
Inbound attachments + parsing
Custom domain inbound.evalbox.ai receives mail with PDF, Word, Excel, and iPhone photos. Tesseract OCR with AI vision fallback for site/scene photos.
Artifacts + email drafts
save_template_draft and propose_email_draft tools with in-thread artifact cards and owner-only Approve & send.
Platform admin + brand polish
/admin observability (traces, LLM usage), branded team invite email, static OG image, Terms/Privacy pages.
Agent Phone
1960s mission-control iPhone PWA backed by Hermes on a Hetzner VPS.
Daily-driver UX on iPhone
Instant tab switches (Research / Build / Invest), session restore from IndexedDB, and breadcrumb context instead of heavy telemetry tiles.
Async Send in all modes
Hermes jobs return 202 immediately — no more Connection lost during long Build/Invest runs. Push notification when the reply is ready.
Transcript copy + search
One-tap Copy on Hermes messages, device search (Now) plus Hermes FTS history search via VPS sidecar.
Hermes VPS integration
Passkey auth, SESSION_BRIEF from evalbox on boot, mode-specific SKILL.md routing, and structured diag logging for field debugging.
$ echo "what's next?"
- PocketForeman — Stripe subscriptions, push notifications
- Field Pipeline — sandbox artifact download, GitHub Action, CSV/Parquet helpers
- AMP — public docs, Slack Connect (Phase 4)
- EB Automate — account settings, Stripe portal, external multi-participant email
- Agent Phone — three-mode integration pass, HEIC file uploads