TECHNOLOGY / FOR EVALUATORS
Sovereignty enforced by topology, not by promise.
A technical brief for ML and security evaluators. These are the engineering principles behind every system Sumedh builds — shown here in DRISHTI, our first platform. Every claim below is a property of how the system is built.
Sovereignty Architecture
DRISHTI's critical path — ingestion, embedding, clustering, post-engine analysis, GIS, and primary inference — executes entirely within the deployment perimeter. There is no outbound dependency required for the system to function. Map tiles are pre-cached; language models run locally; storage is local. The only optional external path is the Groq inference fallback, which is absent by configuration on air-gapped installations. Because no data crosses a national boundary, no foreign data-protection regime has jurisdiction over it.
CRITICAL PATH · WITHIN PERIMETER
Llama 3.1 / Ollama inference and MapLibre + Survey of India GIS attach locally.
Language Stack — MuRIL
DRISHTI embeds text using MuRIL (Multilingual Representations for Indian Languages). English, Hindi in Devanagari, and code-switched Hinglish are represented in a single shared space with no intermediate translation step. The operational consequence: a Hinglish narrative is understood as written, not lossily translated to English first — preserving the code-switching that adversarial messaging often relies on.
Clustering Pipeline — UMAP / HDBSCAN
High-dimensional MuRIL embeddings are reduced to 50 dimensions with UMAP, preserving local and global structure, then clustered with HDBSCAN. HDBSCAN is density-based: it discovers an unknown number of narratives and explicitly labels noise rather than forcing every point into a group. No cluster count is pre-specified and no keyword seed is required — which is what makes detection genuinely unsupervised.
Post-Engine Analysis — PEA Layers
PEA operates on 72-hour sliding windows. PEA-0 anchors integrity with SHA-256 hashing. PEA-1 characterises each narrative. PEA-2 tracks mutation as framing evolves. PEA-3 flags statistically significant spikes at z ≥ 2.0. PEA-4 produces a ranked, exportable assessment with an unbroken, hash-verified chain of custody — the property that makes outputs defensible rather than merely indicative.
PEA-3 · 72h SLIDING WINDOW · z ≥ 2.0
SHA-256 9f3a7c2e…c1d7 · cluster_hash be20af11…7740 · custody ✓ unbroken
Inference & GIS — local reasoning, local maps.
Primary inference runs on Llama 3.1 via Ollama inside the perimeter. The geospatial layer uses MapLibre GL over pre-cached Survey of India tiles, so mapping never queries an external service. Both are designed to satisfy the same constraint: nothing essential leaves the box.