The Pipeline

From raw messages to structured events

OSFeed is not a message aggregator. It's an intelligence engine that automatically collects, translates, classifies, deduplicates, and enriches geopolitical events from dozens of multilingual Telegram channels — in real time.

Step 01·Telegram User Account API — not bots

Real-Time Collection

OSFeed uses the Telegram User Account API (Telethon) to monitor dozens of public channels in real time. Unlike bots, the User API gives full access to all channel content — media, documents, reactions, and edit history. New messages are captured the moment they're published, with automatic rate-limit handling and session persistence.

Channels are grouped by topic: Ukraine-Russia Conflict, Middle East Tensions, Sahel & West Africa, USA & Global Power Shifts. Each topic covers 10–20 curated channels selected for their reliability and coverage.

Step 02·Double translation pipeline

Contextual AI Translation

Every message goes through a two-stage translation process. First, a canonical English translation is generated — this is the shared, reusable version used for event detection, embeddings, and search. Second, if the user's display language differs from English, a second translation is generated in that language and cached.

The translation prompt includes a geopolitical glossary: military jargon, regional acronyms (IDF, FSB, GRU, PMC Wagner), place names, and context-specific terminology. This ensures 'обстрел' becomes 'shelling' not 'shot', and 'منطقة عازلة' becomes 'buffer zone' not 'isolated zone'.

Step 03·AI classification — event vs. commentary

Event Detection

Not every Telegram message is an event. Most messages are commentary, opinions, reposts, or context. OSFeed classifies each translated message using an LLM: does this message report a new fact that happened in the real world (an explosion, a statement, a movement), or is it secondary content?

Only messages classified as event-bearing proceed to the deduplication and enrichment pipeline. This filters out noise and ensures the event panel shows only actionable intelligence.

Step 04·Vector similarity — one event, many sources

Semantic Deduplication

Each event-bearing message is embedded using OpenAI's text-embedding-3-small model. The embedding is compared against existing event centroids using cosine similarity. If similarity exceeds 0.82, the message is linked to the existing event as an additional source. If it falls in a grey zone (0.65–0.82), an LLM decides whether it's the same event.

This is OSFeed's core innovation. A drone strike reported by 14 different channels — in Russian, Ukrainian, Arabic, English — becomes one event with 14 sources, not 14 separate events. Analysts see the corroboration, not the repetition.

Step 05·Events grow over time

Continuous Enrichment

When a new message matches an existing event and adds new information (a casualty count, a diplomatic reaction, a geographic detail), the event summary is updated. The event centroid is recalculated, and the new source is added to the provenance trail. Events aren't static snapshots — they evolve as more reports arrive.

Each event card displays a source count that increments in real time. Analysts can click into any event to see the full timeline: when it was first detected, which sources reported it, and exactly what each source contributed.

Step 06·The split panel — feed vs. events

Structured Intelligence

The final interface shows two panels side by side. Left: the raw feed (chronological messages, translatable, filterable by topic). Right: the event panel (structured events, sorted by latest activity, with source counts and summaries). The contrast is immediate — left is the chaos, right is the intelligence.

Event cards show title, summary, source count, last enrichment time, and topic badge. Clicking an event opens a detail view with all source messages, their original language, the channel name, and a timeline of enrichment. A 'Merge' button lets analysts manually fuse two events if the automated deduplication misses a match.

See it in action

OSFeed is in private beta. Join the waitlist to be among the first analysts to access the platform.

Join the Beta