Skip to content
Matthew Oshin
Back to entrepreneurship
Recent build

Sonar Media

Real-time media monitoring you build in plain English. Describe what to watch and AI assembles the agentic workflow that watches it for you.

Overview

Investor-relations and communications teams drown in signal: a story that moves the stock can break on SEC EDGAR, a newswire, a regulator's feed, or social, and legacy monitoring makes you hand-write boolean queries to catch it. Sonar Media flips that. You describe what you care about in a sentence, an LLM resolves it into a validated monitor spec, and you dry-run it against the last 48 hours before it ever fires. Matches arrive tagged with the source, a sentiment read, the terms that hit, and a one-line summary of why it matters, gated by severity and capped so they never flood you. I built this as the media-intelligence layer of an enterprise platform: thousands of sources, an AI relevance gate to kill noise, and per-workspace AI-spend caps so a monitoring run never becomes a surprise bill.

What it involved

  • A plain-English monitor builder: an LLM turns a sentence into a validated, schema-checked spec (entities, keywords, sources, cadence, delivery), with no boolean syntax.
  • A 48-hour dry run that replays real history, so a monitor earns trust before a single alert goes out.
  • A cheap AI relevance gate on every candidate match, so delivered alerts are the real story and not every keyword hit.
  • Severity gating and per-monitor flood caps: instant for the urgent, digest for the rest, with AI spend tracked per model and capped per workspace.
  • Source-grade coverage across SEC EDGAR, the wires, regulators, social, and cyber feeds, parsed in near real time.

Stack

  • Next.js 16
  • React 19
  • TypeScript
  • Tailwind v4
  • Claude (Opus, Sonnet, Haiku)
  • Zod
  • Async workers
  • PostgreSQL