Built for the full lifecycle
From ingestion to a conversation. Every layer is designed so an analyst — not just an algorithm — retains judgment over what gets surfaced.
1. Ingestion — pluggable, rate-limited, cost-tracked
Independently pluggable source connectors: X/Twitter, YouTube, GDELT news, Meta Ad Library, NewsData.io regional press, and CSV upload — each with its own retry/backoff and per-source legal-clearance flag. A connector cannot go live in production without that flag being set by an authorized admin.
- X (Twitter) — search + engagement metrics
- YouTube Data API — channel comments
- GDELT 2.0 + NewsData.io — Indian regional press
- Meta Ad Library — political ad transparency
- CSV / manual upload for client-supplied data
- Reddit, Telegram, RSS packs — roadmap, legal-gated
2. Processing — dedup, language ID, sentiment, entity linking
Every item is deduplicated, language-identified (including Hinglish/code-mixed detection), entity-linked to tracked candidates/parties/constituencies, and scored by a real multilingual sentiment model — with sarcasm flagged separately, since inversion is a known failure mode for Indian social content.
- Cross-source dedup on near-duplicate text
- Multilingual sentiment (positive/neutral/negative/mixed)
- Sarcasm heuristic, flagged not auto-corrected
- Bot/inauthentic-activity scoring — visible, never silently excluded
- Narrative lineage clustering — social → news → social
3. Nova — the conversational layer
Ask natural-language questions. Nova retrieves the relevant rollups and sample posts, grounds every claim in that data, and cites counts, date ranges, and sample posts — refusing to answer when the data doesn't support a confident response.
- "Why did sentiment on [issue] drop in [constituency] this week?"
- "Compare candidate vs party perception over the last 30 days."
- "Show the top narratives driving negative sentiment, with source lineage."
4. Applicability engine — one config, any entity
A workspace composes entity type (product / brand / candidate / party / issue), geography, lifecycle stage, language scope, and sensitivity tier into a single object — so a party, its candidates, and key opposition figures can be tracked in parallel under one org.