DRISHTI / THE FIRST PLATFORM
One pipeline from open signal to court-ready assessment.
DRISHTI is Sumedh's first sovereign platform — narrative intelligence for the cognitive domain. Six modules, one Docker Compose deployment, zero external dependencies on the critical path. Designed to be operated by an intelligence cell inside an air-gapped network — not consumed as a service.
Ingestion — open-source signal, normalised.
Collects and normalises open-source text signal into a unified store: pgvector for embeddings, MinIO for raw artefacts. Each artefact is hashed on entry, establishing the first link in its chain of custody.
Narrative Engine — meaning-first clustering, no keywords.
Text is embedded via MuRIL — natively handling English, Hindi (Devanagari), and code-switched Hinglish with no translation layer. Embeddings are reduced to 50 dimensions through UMAP and clustered with HDBSCAN, which forms narratives organically and leaves genuine noise unclustered. The operational consequence: analysts see emergent narratives, not a keyword feed they had to configure in advance.
Post-Engine Analysis — five layers from cluster to evidence.
- PEA-0 — Baseline integrity: SHA-256 hashing and provenance anchoring of every artefact and cluster.
- PEA-1 — Narrative characterisation: theme, language mix, and structure of each cluster.
- PEA-2 — Mutation tracking: how a narrative's framing shifts as it propagates.
- PEA-3 — Spike detection: statistically significant surges flagged at z ≥ 2.0 over 72-hour sliding windows.
- PEA-4 — Assessment & export: ranked narratives with a complete, hash-verified chain of custody suitable for evidentiary use.
Geospatial Layer — narratives on the ground, no external map called.
MapLibre GL renders narrative geography over pre-cached Survey of India tiles. Because tiles are local, the GIS layer functions fully air-gapped and makes no call to any foreign mapping API.
Sovereign Inference — local reasoning, with a controlled fallback.
Summarisation and reasoning run on a local Llama 3.1 instance via Ollama. Where an environment permits connectivity, a Groq API path is available as an online fallback — never required, and never present on the air-gapped deployment.
Analyst Interface — ranked narratives, full provenance, one screen.
A React 18 console presents clustered narratives, mutation timelines, spike alerts, and the geospatial layer — each item traceable back to its source artefacts and hashes.
Every box runs inside your perimeter.
CRITICAL PATH · WITHIN PERIMETER
GIS and inference run as local side services.
Deployment options.
On-premise, air-gapped
Single docker compose up on customer hardware. No outbound connectivity. Recommended for defence and intelligence use.
- No external API on critical path
- Llama 3.1 local inference only
- Survey of India tiles pre-cached
On-premise, connected
Same deployment with the optional online inference fallback enabled, for environments where controlled connectivity is permitted.
- Groq fallback available
- Operator-controlled egress
Sovereign cloud
A managed deployment for policy bodies, election administration, and strategic-analysis users who do not require air-gapping.
- Lighter footprint
- Same narrative engine
What DRISHTI is not.
It is not a social-media monitoring dashboard, not a keyword alerting tool, and not a hosted service that ingests your data on someone else's infrastructure. It is owned infrastructure for narrative intelligence — the first platform Sumedh has built on its sovereign cognitive foundation, with more to follow.