Imara

Imara Story

Inspiration Online gender-based violence is a growing epidemic, and victims often lack accessible, secure, and discreet ways to seek help or gather admissible evidence. Our inspiration for Imara was to build a "Zero-UI" Digital Bodyguard—a system that requires no app downloads and hides in plain sight with camouflaged interfaces that look like generic news services. We wanted to create an indestructible, multimodal support system tailored specifically for Africans, capable of understanding local dialects and navigating complex jurisdictions in real-time.

What it does Imara is an ecosystem powered by a specialized "Bodyguard Hive" network of 7 micro-agents, designed to tackle the Live Agents and Creative Storyteller tracks. Users can interact with Imara seamlessly via Telegram, Meta (Messenger/Instagram), or a Web PWA.

The 7-agent Hive includes:

• Sentinel: Enforces safety policies and detects violations.
• Linguist: Specializes in African dialects (like Pidgin and Swahili) and tonal analysis.
• Visionary: Uses advanced multimodal OCR and visual threat detection to analyze screenshots of abuse.
• Navigator: Handles high-precision jurisdictional routing across 54 African countries.
• Forensic: Acts as a chain-of-custody expert, generating legally admissible SHA-256 cryptographic evidence hashes.
• Messenger: Automates the drafting of verified mixed-media alerts for support partners.
• Counselor: The empathetic "Supportive Sister" persona for user-facing, real-time dialogue.

How we built it We built Imara using a modern, highly optimized 2026 tech stack designed to run entirely on a Google Cloud 1GB RAM always-free VM.

• Backend & Infrastructure: We used Python and Django 6.0 with Uvicorn (ASGI) for native async support. By deploying on the Google Cloud 1GB VM, we had to be incredibly resource-efficient, utilizing SQLite in WAL mode and Django Native Tasks to eliminate the need for heavy external dependencies. We also leverage GCP HTTPS offloading.
• AI & Orchestration: The core brain is orchestrated using the ADK (Agent Development Kit). The ADK allowed us to seamlessly wire our 7-agent "Bodyguard Hive" using a "One Agent, One Tool" pattern. Gemini 2.5 Pro serves as our primary multimodal engine and Visionary agent, analyzing screenshots of online abuse and interpreting visual elements securely. We implemented Semantic Disk Caching for instant reasoning on recurring location and policy queries, drastically reducing memory overhead.

Challenges we ran into One of the biggest challenges was engineering a complex 7-agent architecture via the ADK that could run smoothly on a highly constrained 1GB RAM Google Cloud instance without sacrificing speed or reliability. We had to heavily optimize our background task processing and pass standardized ContextBundle objects between the ADK agents to prevent memory bloat.

Accomplishments that we're proud of We successfully built a production-ready, forensic-grade multimodal AI pipeline that operates entirely invisibly to the attacker. Achieving native async orchestration in Django while strictly adhering to the memory limits of the Google Cloud free tier is a massive technical win for accessibility.

What we learned We learned how to effectively leverage the ADK to isolate AI model responsibilities, reducing token usage and improving response times. We also learned how powerful Gemini's native multimodal capabilities are when interpreting localized visual threats across different social media platforms.

What's next for Imara Next, we plan to onboard more grassroots partner organizations across the continent, expanding our Agent Pro custom personas, and integrating deeper interleaved audio capabilities for real-time voice interventions.

Built With

• adk
• cloudflare-r2
• django
• gemini-2.5-pro
• gemini-multimodal
• google-cloud
• google-compute-engine
• litellm
• python
• sqlite
• telegram
• uvicorn