Most hackathon projects die on the demo table. They work once, in the right browser, with the right seed data, connected to the right Wi-Fi. careerio was one of them — built in 48 hours, deployed at 4am, demoed to a panel of judges who had never heard of the protocol it was built on. It won.

But more than winning, building careerio taught me something about what open protocols can actually do when you give developers the keys — and about what they still can't do.

What the beckn protocol actually is

beckn is an open, decentralized protocol for digital commerce. Think of it the way you think of HTTP — a standard that lets different systems talk to each other without any one company owning the connection. On the web, HTTP lets any browser reach any server. With beckn, any consumer-facing app can reach any seller's backend without a marketplace intermediary.

In the jobs context, this means a job seeker using careerio could theoretically search across LinkedIn, Indeed, Naukri, and any other platform that implements the protocol — from a single interface that none of those platforms own. The protocol handles the search, discovery, and application flow. The platforms just serve their data.

The 48-hour build

The stack: React + TypeScript on the frontend, a Node.js middleware layer to handle beckn message formatting, and OpenAI's API for two things — resume parsing and job matching scores. The frontend was the easy part. The beckn integration was not.

beckn's documentation is structured for architects. If you are a developer with 48 hours and a deadline, you will feel this immediately.

The protocol uses a specific JSON structure for every message type — search, select, init, confirm. Each message has a context object with transaction IDs, timestamps, and BAP/BPP identifiers that have to be correct or the network rejects you silently. We spent roughly 10 of our 48 hours just getting the first successful search response.

Where the AI actually helped

We used OpenAI in two places. First, resume parsing: users upload a PDF and the API extracts skills, experience, and preferred roles as structured JSON. This powers the initial job search query without asking the user to fill out a form.

Second, match scoring: every job result coming back from the beckn network gets scored against the user's resume. Not a keyword match — a semantic comparison using embeddings. A React developer's resume matches 'Senior Frontend Engineer' even if the words don't overlap exactly.

• Resume parsing with GPT-4 Turbo: accurate, fast, handles any formatting
• Semantic job matching via embeddings: better signal than keyword search
• Generated cover letter drafts per application — the feature users wanted most
• Fallback gracefully when the beckn network had no results: showed OpenAI-sourced suggestions instead

What I'd do differently

The beckn network coverage for jobs is still thin. The protocol is technically sound but the adoption isn't there yet — most searches returned 0–3 results because very few job platforms have implemented it. We papered over this with AI fallbacks, which worked for the demo but would fail in production at scale.

If I were building careerio for real, I'd start with a specific vertical — government jobs in India, for instance, where ONDC and beckn adoption is being mandated — rather than trying to aggregate everything. Narrow focus, deep integration, real data. The ambient potential of an open protocol is worth nothing if nobody has plugged in.