• Written by Rob Kleiman

OliveWing: Building the Intelligence Layer a Vibe-Coded MVP Couldn't Deliver

TL;DR

  • OliveWing turns online conversation into audience intelligence for political and advocacy clients. Its fast-built MVP kept retrieving the wrong content with high confidence.
  • The fix was a reframe: it wasn’t a search tool, it was a structured knowledge system with retrieval on top. That meant a graph database (Neo4j), source-aware ingestion and a schema that knows who is speaking, toward whom, and how they feel.
  • The work helped get OliveWing included in Substack’s midterms roundup.

The company

OliveWing is an audience intelligence platform for political and advocacy clients. The promise: take in large volumes of Reddit threads, Facebook groups, podcasts and news, and surface structured insight about how real people think, speak and feel. Clients ranged from political campaigns needing message-resonance scores to super PACs tracking voter factions across a state.

  • Industry: political tech and AI
  • Stage: early product, first clients
  • Engagement: architecture and build

The problem

OliveWing had a real idea and a working product. But as they pushed it against real client data, the cracks became impossible to ignore.

The app had been built fast, vibe-coded into existence. Its pipeline processed all content the same way regardless of source, and its retrieval treated semantic similarity as a stand-in for meaning. In practice, the product regularly surfaced the wrong content with high confidence: a sarcastic Reddit comment about a senator retrieved as support for that senator, or a post that mentioned a politician retrieved as being about that politician when it was really about something else.

Underneath the interface, the system had no idea who was speaking, what they were speaking toward, or how different factions in the same conversation might be responding in opposite directions.

What we did

Started from the foundation

The engagement started with a working session with the founder: not a brief, not a deck, a whiteboard conversation about what the product needed to do and where the current build couldn’t get there. What came out of it was a reframe. OliveWing wasn’t a search tool. It was a structured knowledge system that needed retrieval on top of it. That changed how the whole architecture got rethought.

1. Chose a graph database as the backbone

Flat vector retrieval was the root of most of the problems: it had no concept of relationships, direction or context. We evaluated graph databases as the backbone, testing whether the shape of the data (who said what, in response to what, toward whom) was better modeled as a graph than as embeddings alone. It was, and that decision shaped everything downstream. The build used Neo4j, added to OliveWing’s Replit app.

2. Made ingestion source-aware

The pipeline had treated a Reddit thread the same as a podcast transcript. We redesigned ingestion so each source type gets its own chunking thresholds, threading logic and normalization rules: community discussion, long-form audio and editorial content each behave differently. Community content came in through Apify actors for Facebook posts and Facebook groups. This alone meaningfully improved what the system could retrieve, and why.

3. Designed the metadata schema

The most consequential work was deciding which structured labels every chunk of content needed: stance toward the subject, speaker direction, faction identity, and whether an entity is merely mentioned or actually the topic. These weren’t features. They were the difference between a system that retrieves content and one that understands it. We designed the schema and a four-pass enrichment process that attaches these labels as content comes in.

4. Picked the toolchain to fit the business

With the architecture clear, we mapped each layer (discovery, fetching, normalization, orchestration, enrichment, retrieval, observability) to a build, buy or integrate decision. The goal was to build only what was genuinely OliveWing’s own intellectual property and integrate the rest with tools that could scale as the client list grew across campaigns, PACs and nonprofits.

What changed

  • A system that knows context: the high-impact conversations OliveWing sources and curates are now mapped to candidate coverage and to the people running, instead of being matched on keywords.
  • Recognition: the work helped get OliveWing included in Substack’s midterms roundup.

What I took away

Vibe coding gets you a first version fast, which is exactly why the hard part has moved. The question isn’t “can we build it?” anymore; it’s “what does this actually need to understand?” Answering that took a whiteboard and a founder, not more code.

References