Devi is an AI social-media monitoring and lead-generation tool. It watches keywords across nine platforms — Facebook groups (public and private), LinkedIn, X, Reddit, Nextdoor, WhatsApp, Bluesky, Threads and Telegram — detects posts that signal buying intent, complaints or questions, and then uses AI to draft personalised comments and outreach. The pitch: “More Organic Leads, Less Effort”, replacing $100-a-lead paid ads with 24/7 automated listening, from $19/month.
What makes Devi architecturally interesting is where the work happens. Most of it runs in the user’s own browser through a Chrome extension, not on a central server. Serverless is used surgically — for one job only: the AI intent detection. That split is the whole story.
The browser does the heavy lifting
Devi’s Chrome extension runs inside the user’s logged-in browser session, and that’s deliberate. From there it handles:
- Scanning & monitoring. The extension reads the social feeds and groups the user already has access to, checking them for the user’s keywords. Because it runs client-side using the user’s own session, there are no platform API limits to fight and no credentials to store server-side.
- Authentication & user management. Sign-in and account state are managed in the browser, keeping the moving parts close to the user.
- Privacy by design. Devi notably stores no scraped social data on its own servers. Since the scanning happens locally and only the relevant snippet is ever sent out for analysis, that posture is easy to honour.
This client-heavy design means Devi has no fleet of servers polling nine platforms around the clock for every customer — which would be expensive, rate-limited and a privacy liability. The browser each user already runs is the monitoring infrastructure.
Where serverless comes in
When the extension finds a candidate post, it can’t decide on its own whether the post is a real buying signal, a complaint or just noise — that takes a language model. So the extension makes a short call to a serverless function that runs the AI intent detection, and gets back a verdict (and, where relevant, a suggested reply). That’s the one piece that belongs in the cloud, and serverless fits it perfectly:
- Pay-per-call inference. A function spins up only when there’s a real candidate post to judge, runs the model, and scales straight back to zero. Cost tracks actual signal volume, not wall-clock time.
- Naturally bursty. Intent checks arrive in unpredictable spikes as users browse. Concurrent function invocations absorb the spikes without any provisioning.
- Stateless and isolated. Each inference call is independent — no session to keep warm, nothing persisted — which keeps the privacy story clean and the architecture simple.
The payoff
By keeping scanning, auth and user management in the browser and reserving serverless purely for AI intent detection, Devi gets the best of both worlds: a monitoring layer that costs the company almost nothing to run and respects user privacy, and AI inference that scales on demand and is billed only when it actually fires. It’s a lean split that lets Devi sell an always-on service from $19/month while still running real LLM analysis under the hood.
Devi is a great illustration that serverless doesn’t have to run your whole product — just the part that benefits most. Push what you can to the client, and use serverless surgically for the on-demand, compute-heavy piece. If you’re building extension-plus-AI tooling like this, our serverless specialists design exactly these client-and-serverless splits. See Devi for yourself at ddevi.com.