Watchdog Strategist

Inspiration Every startup says "keep an eye on the competition" — but nobody actually has time to open five pricing pages and three changelogs every morning. So a competitor's price cut or new feature gets noticed a week late, and by the time someone writes the customer email or the social post, the moment has passed. We wanted an agent that does the boring watching and the first draft of the response, so a small team can react like a big one — without hiring anyone to stare at competitors all day.

What it does Watchdog Strategist is an autonomous competitor-intelligence agent. On a schedule you choose, it runs in the background and, for every competitor you track:

Watches their pricing page, changelog, news, and reviews. Detects meaningful changes by diffing against the last saved snapshot (and ignores page noise). Analyzes each real change with Gemini 3.5 Flash — is this a threat or an opportunity, how severe is it, and what's the estimated market impact (0–100, with the model's reasoning shown as a chart). Drafts your response — a ready-to-review customer email and a social post, written in your product's voice, positioning you against that exact competitor move. Lets a human decide. The dashboard shows briefings Outlook-style (list + reading pane). You Approve, Edit the drafts inline, or Dismiss. Then it helps you act, keeping you in control the whole way:

Send to SmartMail turns the approved email into a branded HTML campaign draft in your own email app — it never auto-sends; you hit send. Share opens X / LinkedIn / Facebook / Instagram with the post text (and a branded 1080×1080 social card) ready to publish. On scheduled runs, important briefings (high severity or high impact) email the team a digest and fire an Android push notification, so you find out the moment it matters — even away from your desk. There's also an admin panel to manage competitors, your product profile, email branding, recipients, and the check interval — all stored in Firestore, no code change needed — plus an Overview page with trend charts and a mobile app (Android WebView + push).

How we built it Agent brain: Gemini 3.5 Flash via the Google GenAI SDK, as a two-step Analyst → Strategist chain behind a small LLM boundary, so the reasoning is swappable without touching the orchestration. Backend: FastAPI on Cloud Run. Autonomy: Cloud Scheduler wakes POST /run on the configured interval — the agent runs with no human clicking anything. State: Firestore stores page snapshots, briefings, the competitor list, settings, and device tokens. Watching: httpx + BeautifulSoup for fetching and clean-text extraction; Google News RSS for news and review signals. Notifications: the team digest and campaign drafts go to the user's own email app over a small HTTP API; Firebase Cloud Messaging for Android push. Mobile: a Kotlin WebView wrapper app that registers an FCM token. Quality: 115 automated tests (pytest) across the detector, watcher, agents, storage, orchestrator, impact math, digest, push, and every route. Data sources Competitor pricing and changelog pages (fetched live). Google News RSS for competitor news and for review / complaint / alternative signals. All derived state (snapshots, briefings, settings) lives in Firestore. Challenges we ran into Model region: Gemini 3.5 Flash returned 404 in some regions — we had to serve it from GOOGLE_CLOUD_LOCATION=global. Page noise: dynamic pages "change" on every fetch. We added a noteworthy flag so the agent suppresses non-changes instead of spamming junk briefings. Keeping the human in control: it's tempting to auto-send. We deliberately made every outbound action a draft the human approves — the email lands as a draft, the social share opens the composer pre-filled. The agent does the work; the person makes the call. Resilience: competitor sites and news feeds occasionally block or return 503; the watch cycle degrades gracefully and runs on the sources that succeed. Accomplishments that we're proud of A genuinely autonomous loop (schedule → watch → detect → analyze → draft → notify) that produces something a busy founder would actually send. An honest human-in-the-loop design — no fake "auto-blast to thousands." End-to-end on Google Cloud (Cloud Run + Firestore + Cloud Scheduler + FCM) with a real mobile app and 115 passing tests. What we learned The hard part of an "agent" isn't calling the model — it's deciding what's worth surfacing (the detector + noteworthy gate) and where to stop and ask a human (approve before send). Those two boundaries are what make it feel trustworthy instead of noisy. A tight prompt contract (structured analysis with severity + impact) makes the output consistent enough to render as charts and drafts, not just text. What's next for Watchdog Strategist Real one-click publishing via the social platforms' APIs (today it's human-in-the-loop share). Multi-tenant "agency" mode (watch competitors on behalf of many clients). More verticals beyond SaaS (retail sale-pages, and more).

For judges — test access The public read-only dashboard is open at the live URL. To try the actions (Approve / Edit / Send / admin panel), open /admin and log in:

URL: https://sentinel-480188517080.us-central1.run.app/admin Password: demo1234 To watch the agent run a cycle, use the Run cycle button on the dashboard header (or POST /run?manual=1).