DEV Community

Cover image for Building SARA: A Sovereign Multi-Agent AI System on Google Cloud to Eliminate Extortion and Break Judicial Impunity
carlosmba12Peru
carlosmba12Peru

Posted on Edited on

Building SARA: A Sovereign Multi-Agent AI System on Google Cloud to Eliminate Extortion and Break Judicial Impunity

#ai #python #googlecloud #security

Declaration of Participation: I created this piece of content for the purposes of entering the "All Things Agentic Hackathon" organized by Google Cloud and Devpost.

Building SARA: A Sovereign Multi-Agent AI System on Google Cloud to Eliminate Extortion and Break Judicial Impunity

Extortion, violent racketeering ("cobro de cupos"), and predatory lending ("gota a gota") represent one of the most critical public safety emergencies across developing nations. In Peru alone, official records document over 200,000 real extortion attacks annually, yet over 85% of victims remain in total silence due to mortal terror of identity leaks and lethal retaliation.

Official judicial reports from the Peruvian Institute of Economics (IPE), Public Prosecution (MPFN), and Judiciary Flagrancy Courts (2026) expose a devastating institutional bottleneck: only ~240 cases reach sentencing per year, leaving a staggering 99.88% systemic impunity rate.

To dismantle this crisis, I designed and developed SARA (Sistema Autónomo de Respuesta Anti-Extorsión) for the Google Cloud & Devpost #AllThingsAgenticHackathon.

🌐 Live Production Demo & Links

🏛️ 1. Parallel Multi-Agent Swarm Architecture

SARA is an enterprise-grade, serverless multi-agent operating system deployed on Google Cloud Run (us-central1) and powered by the Google Agent Development Kit (ADK).

Instead of a single monolithic LLM prompt, SARA orchestrates a Hierarchical-Parallel Swarm of 17 specialized agents executing 4 cognitive branches simultaneously in under 1.8 seconds:

  • Branch 0 (Kallpa / Amparo): Empathetic containment and code-switching in 5 indigenous languages (Quechua, Shipibo-Konibo, Aimara, Asháninka, Awajún) + Spanish & English.
  • Branch 1 (Forensics & PIDE Agent): Cross-referencing OSIPTEL-RENTESEG (3h IMEI cut under Law 32303), RENIEC, and multimodal OCR ballistic evidence certification.
  • Branch 2 (Threat Calculator): Deterministic mathematical scoring using Thomas Saaty's Analytic Hierarchy Process (AHP-Saaty) to compute the Extortion Risk Index (IRCE: 70% Certainty / 30% Imminence).
  • Branch 3 (Legal Advisor & Packager): Subsumption under Articles 200 & 200-A of the Criminal Code and immediate digital transmission to the Specialized Anti-Extortion Prosecution Office (FECOR - D.Leg. 1735).

🧠 2. Dual-Brain Cognitive Routing & Zero-PII Cryptography

SARA leverages Google Cloud's foundational models via the Google GenAI Python SDK (google-genai):

  • Gemini 2.5 Flash / Gemini 3.7 Flash: For real-time acoustic triage (<300ms latency), anti-spam filtering (D.S. 020-2020-MTC), and low-latency voice containment (<600ms).
  • Gemini 3.7 Pro Reasoning (thinking_budget=2048): For deep criminal intelligence deduction, multimodal handwritten note OCR, and statutory compliance with zero hallucinations.
  • Sovereign Zero-PII Cryptographic Vault: Citizen identity is isolated at ingestion using HMAC-SHA256 with CSPRNG salt and Cloud KMS (HSM FIPS 140-3), issuing an anonymous Unique Protection Code (CUP) so LLMs never handle real personal data.

👮 3. Human-in-the-Loop Sovereign Governance (Law No. 31814)

In compliance with Peruvian AI Law No. 31814 and ISO/IEC 42001:

  • AI assists, but a verified human police commander retains absolute decision-making power.
  • The officer reviews the dossier, validates tactical measures (24h preventive bank freeze under D.Leg. 1735 / UIF and <3h IMEI cutoff), and signs with their Official Police Digital Token (CIP) into SIDPOL.
  • Result: Response time collapses from 48 hours of manual paperwork to 1.8 seconds of judicial-grade forensic action.

⚠️ Institutional Independence & Legal Disclaimer

SARA is an independent research, software prototype, and demonstrative project developed by Carlos Eduardo Baños Diaz exclusively for the Google Cloud & Devpost #AllThingsAgenticHackathon.

It is not an official platform of the Peruvian National Police (PNP), the Ministry of the Interior (MININTER), or the Public Prosecutor's Office (Ministerio Público).

All references to Peruvian public institutions, statutory frameworks (Legislative Decrees No. 1735, 1731, Law No. 32303, R.M. 152-2026-PCM), and open datasets (SIDPOL, IPE reports) are used strictly as a realistic domain engineering model.

🔒 Synthetic Data Disclosure (Law No. 29733): All citizen names, national IDs, and financial records used in demonstration cases are 100% synthetic and procedurally generated for AI benchmarking.

Top comments (0)