Back to projects

Lomita OS

A supervised AI command center for relationship-driven investment teams—agents do the work, humans stay in control.

L
View Live

Overview

Lomita OS is the product behind Lomita AI: a supervised command center where investment teams manage a fleet of AI agents. The first vertical slice is an outbound email agent that moves a contact from CRM selection through grounded draft generation, human edits or feedback-driven revisions, and an explicit Gmail-approved send—then evaluates the first draft against the exact version that was sent, so the agent measurably improves over time.

Key Features

  • CRM workspace with company, relationship, and next-action context, plus on-demand AI contact enrichment
  • Grounded email drafting with immutable draft versions—every edit and feedback turn is preserved
  • Human-in-the-loop review queue: the model has no send tool, and every send requires explicit approval
  • Writing profile and evidence-backed rules learned from your own sent email history
  • Deterministic evals comparing each first draft to the final sent version

    • Technical Details

    Built with Next.js and the Vercel AI SDK via AI Gateway. Supabase powers the multi-tenant backend with row-level security and encrypted Gmail history. Langfuse handles observability, with AI inputs and outputs omitted from hosted traces.

    Learnings

    Building agents for finance taught me that safety is the product: append-only audit records, AES-256-GCM encryption for mailbox data, server-side compare-and-swap checks on every send, and keeping the model physically unable to send email are what make an agent system professionals can actually trust.