Strategos

Summary

Strategos is a strategic intelligence platform purpose-built for the Bangladesh 2026 national election, combining 300 AI constituency agents, multi-agent synthesis, and a 313-person field reporting network. It pairs historical election results from 1991, 1996, 2001, and 2008 with live ground intelligence to produce daily briefings, victory path projections, and seat-by-seat recommendations. Originally developed for the BNP campaign command, the underlying architecture generalizes to any constituency-based electoral system.

What it is

Strategos is a case study in how to build a campaign-grade intelligence system that operates simultaneously at the national, divisional, and constituency level. The platform deploys one dedicated AI agent per parliamentary seat (300 in total), each conditioned on that constituency's historical voting record across four prior elections, demographic composition, candidate matchups, and the most recent field report from the local operative. A multi-agent layer above this (Analyst, Reporter, Watcher) synthesizes seat-level signals with external news flow into a single daily executive briefing, with threat levels, top opportunities, and prioritized action items.

The human side mirrors the AI architecture: 5 headquarters commanders, 8 divisional commanders, and 300 field agents submit structured reports on visibility, ground activity, sentiment, opposition tactics, and urgent issues, all gated by role-based access control and a 6-hour submission cooldown. GitHub Actions runs external intelligence refreshes four times daily and a full field-plus-news synthesis once daily, with results written to a Supabase backend that powers a React dashboard with executive, divisional, and constituency views. A photocard generation pipeline produces shareable candidate and constituency assets for field distribution. While the original deployment was tuned to Bangladesh's 300-seat parliament and the BNP context, the methodology, agent-per-constituency design, historical baseline integration, hierarchical field network, and automated synthesis cadence, transfers directly to any first-past-the-post or constituency-based electoral environment that needs ground-truth intelligence at scale.

Methodology

• One AI agent per constituency (300 total), each conditioned on local history, demographics, and current candidate matchup
• Three-agent synthesis layer (Analyst, Reporter, Watcher) for daily executive briefings combining field and external signals
• Hierarchical human network of 5 HQ, 8 divisional, and 300 field operatives feeding structured ground reports
• Historical baseline integration using actual results from 1991, 1996, 2001, and 2008 elections per seat
• Divisional dashboards with risk matrices, swing seat analysis, and victory path projections
• Automated cadence via GitHub Actions: external intel four times daily, full synthesis once daily at 6 AM Dhaka time
• Photocard generation pipeline for candidate and constituency assets distributed to field agents
• Role-based access control with bcrypt authentication, row-level security, and submission cooldowns to prevent data poisoning

Data sources

• Historical Bangladesh parliamentary election results, 1991 through 2008, by constituency
• Structured field reports from 300 constituency-level operatives covering visibility, sentiment, and opposition activity
• Constituency demographics including voter rolls, gender split, urbanization, and youth share
• Google News and external media coverage retrieved via Gemini for opposition and sentiment monitoring
• Election Commission gazette data for candidate registration and constituency boundaries

Deliverables when used in engagements

• Daily executive briefing with threat level, top opportunities, and prioritized action items
• Per-constituency intelligence cards with win probability, classification, and seat-specific recommendations
• Divisional dashboards covering victory path, momentum, and swing seat analysis
• Field agent reporting portal with structured submission flow and role-based access
• Candidate and constituency photocards generated programmatically for ground distribution