$→
Open Money Intel
Subscribe
How this is wired

Three loops, one source of truth.

Open Money Intel is a working prototype of three PMM workflows on the same architecture: a competitive-monitoring loop that watches stablecoin rails, a launch pipeline that compresses positioning into a full asset package, and a measurement loop that feeds channel outcomes back into the positioning YAML as concrete edit suggestions.

Walkthrough
Three-minute walkthrough
What this is, how it's wired, and the loop the JD asks for.

System map

DefiLlama Stablesstablecoin supply per chainL2BeatTVS, activity (secondary)Normalizer · /lib/datamap chains → rails; compute 7d/30d ΔDashboard/rails, /briefing, /Weekly crongenerateObject → digestResendweekly emailpositioning/*.ymlsingle source of truth · platform + N productsgenerateObject pipeline6 calls per product · 1 platform call · Zod-validatedLaunchBundletyped assets + meta/launch viewertabs · copy · exportLOOP 1 · MONITORINGLOOP 2 · LAUNCH PIPELINEPMM edits positioning.yml → regenerates → reviews → ships · time-to-market measured in minutes
Continuous · cron-driven

Loop 1 — Monitoring

  1. 01Pull from DefiLlama Stablecoins + L2Beat hourly (Next.js fetch revalidation).
  2. 02Normalize seven payment rails (Polygon Stack, Tron, Solana Pay, Base, Circle, TON, BSC) with computed 7d/30d deltas.
  3. 03Every Monday 08:00 IST, a Vercel cron triggers /api/cron/weekly.
  4. 04Claude (via Vercel AI Gateway) gets the rail data and a Polygon-PMM system prompt. It returns a Zod-validated briefing object with 4 sections.
  5. 05Resend dispatches the digest to subscribers. Each delivery is logged.
On-demand · positioning-driven

Loop 2 — Launch pipeline

  1. 01positioning/*.yml is the source of truth — Polygon Stack and Circle Network are wired side-by-side to prove generality.
  2. 02POST /api/launch/generate parses the YAML through a Zod schema (rejects malformed positioning before any AI call).
  3. 03Per product, six parallel generateObject calls produce one-pager, landing block, 5-email sequence, 10 LinkedIn ads, 3 battlecards, BD talk track.
  4. 04One platform-level call produces the master narrative, bundled pitch, ICP-to-product map, cross-product table.
  5. 05Each call returns a typed object — never free-form text — so the viewer can render it without parsing.
  6. 06Tokens, latency, and estimated cost surface in the observability panel. /launch/debug shows every prompt and schema.
Closes the cycle · YAML-aware

Loop 3 — Measurement

  1. 01Assets fan out to six channels: LinkedIn ads, landing pages, email nurtures, one-pagers, battlecards, BD talk tracks.
  2. 02Each channel reports impressions, CTR, conversion rate, spend, CPL — seeded deterministically from the source hash for demo reproducibility.
  3. 03The reporter surfaces the worst-CPL or lowest-converting asset per product as a finding.
  4. 04Each finding ships with a concrete YAML edit: move a proof point, swap a competitor framing, promote a FAQ into differentiators.
  5. 05PMM applies the edit → regenerates → telemetry seed changes → next iteration. The loop is the product.
  6. 06Live at /telemetry. Production swaps the seeded data for LinkedIn Campaign Manager + GA4 + Salesforce.

Decision principles

Structured output, always

Every AI call returns a Zod-typed object. No regex on prose. Voice violations surface as schema failures, not silent drift.

One positioning source

The YAML is the bottleneck on purpose. Change tone, swap a competitor, edit a differentiator — the whole asset pack regenerates from the same source.

Fallback that's still useful

When the AI key is missing, deterministic templates produce shippable (not generic) drafts. Demos work offline; iteration with Claude is the upgrade.

Velocity surfaced

Latency + token + cost panel. The point of agent infrastructure is that ten asset families generate faster and cheaper than one PMM contractor would draft them.

Honest data

Tron is bigger than Polygon. The dashboard says so. A briefing that hides the truth fails its readers — fintech buyers do the same math.

Composable, not magic

Each loop is a small set of explicit pieces. Anyone on the team can edit the YAML, regenerate, and ship in minutes.

Tech stack

›Next.js 16 · App Router · Turbopack
›Vercel — Functions, cron, AI Gateway
›Claude Sonnet 4.6 via AI Gateway
›Zod schemas on every AI boundary
›Neon Postgres + Drizzle ORM
›Resend for transactional email
›DefiLlama Stablecoins · L2Beat APIs
›Tailwind 4 · Recharts
›In-memory dev fallback for everything

What I'd build next

  1. Linear/Notion integration.When a product update lands in Linear, auto-draft the launch pack against the new positioning and open a Notion doc for review.
  2. Live competitor-page diffing.Weekly crawl of named-competitor homepages and pricing pages; alert when copy or value-prop shifts.
  3. Slack approvals.Generated assets post to a Slack channel with thumbs-up gating before going live to BD.
  4. A/B-able asset variants.Each landing block generates with three angles; the one with highest LinkedIn ad CTR gets promoted to the live site.
Why this exists

Open Money Intel is a demo for the Polygon Labs Product Marketing Manager role. The JD describes the work directly: positioning-as-source-of-truth, AI workflows that run without a human in the loop, and an honest read of where Polygon is competitively winning and losing. That's what this site is.