Neural 360

Neural 360 unifies every customer signal—sales, support, usage, sentiment—into one unified profile, turning scattered data into real-time, explainable recommendations that grow revenue and cut churn

Comment

Inspiration

  • Every enterprise faces the same problem: one customer leaves traces across a dozen disconnected systems — sales, support tickets, reviews, social, product usage, contracts.
  • Nobody can see the whole person; teams spend weeks stitching together a partial picture that's stale by the time it lands.
  • We wanted to collapse that fragmentation into one living, explainable view of each customer.
  • The bet behind Neural 360: instead of building yet another warehouse, make MongoDB Atlas the corporate master datastore an AI can read directly.
  • Goal: let anyone ask a plain-English question and get a trustworthy answer in seconds.

What it does

  • Unifies every customer signal into one AI-powered behavioral profile.
  • Ingests structured records, unstructured conversations, and documents into a single MongoDB Atlas collection.
  • Answers natural-language questions with a fused dashboard covering:
    • Health, churn risk, growth, and adoption scores
    • Spend trajectory and sales prediction
    • Sentiment analysis
    • Recommended next actions
    • A behavioral timeline
  • Every figure is traceable back to its source.

How we built it

  • Made MongoDB Atlas the single master datastore, normalizing 20+ formats into one unified JSON schema.
  • Structured rows, text, and 384-dimensional embeddings live as siblings in the same collection.
  • Atlas does triple duty — operational store, Atlas Search for keyword (BM25) retrieval, and native $vectorSearch — collapsing three databases into one.
  • The MongoDB MCP Server makes the store directly addressable by an LLM.
  • A meta-prompting layer inspects the live schema, classifies intent, and writes its own aggregation pipeline.
  • Retrieval fires structured, keyword, and vector queries in parallel, fuses them with reciprocal rank fusion, and reranks with a cross-encoder.
  • A scoring layer rolls signals into explainable customer scores.
  • Autonomous agents handle PII detection, data protection, and quality continuously alongside human stewards.

Challenges we ran into

  • Cross-customer data leakage — fusing tickets, contracts, and reviews could mix records; we made a customer scope mandatory on every document and injected the filter at query-generation time rather than trusting the model.
  • Vector recall on long documents — the embedding window was far smaller than our text cap; we fixed it by chunking before embedding and linking chunks back to their parent.
  • BM25 scaling — an in-memory index didn't scale and went stale on new ingests, so we moved keyword search onto native Atlas Search.
  • Mixed-modality fusion — reconciling exact aggregations with fuzzy $vectorSearch hits in one ranked list forced an explicit intent-routing and tie-break policy.

Accomplishments that we're proud of

  • Went from raw files to a working platform in three weeks.
  • Unified 20+ source formats into one Atlas collection.
  • Stood up a parallel ensemble retriever — structured, Atlas Search, and vector fused by RRF — that beats any single method on recall.
  • Shipped a dashboard that reconstructs a full customer profile from just a name.
  • Every answer is explainable and scoped to a single customer — no black-box numbers, no leakage across boundaries.
  • No separate query API to maintain, thanks to the MCP Server.

What we learned

  • The hard part of customer intelligence isn't the model — it's the data shape.
  • Once every format shares one MongoDB schema, retrieval and reasoning get dramatically simpler, and a new source onboards in hours instead of a quarter.
  • No single retrieval method is enough.
  • Letting Atlas serve as operational store, search engine, and vector index removes enormous integration glue.
  • Cheap models should handle volume while large models handle nuance.
  • Governance — scope, purpose, and PII handling — has to be designed into the document schema from day one, not bolted on later.

What's next for Neural 360

  • Harden the tenant-scoping and PII model for production data, with mandatory filter injection and PII classification baked into the schema.
  • Expand the autonomous agent workforce for deeper data-quality and protection coverage.
  • Close the recommendation loop with feedback and realized-impact tracking.
  • Add more connectors into the unified collection.
  • Enable richer agentic reasoning over Atlas via the MCP Server.
  • Tighten integration with the workflows where teams actually act on these insights.

Built With

Share this project:

Updates