PRAKHARAI SYSTEMS THAT DEFEND PEOPLE
PS©2026
← Back

02 · Autonomous fraud-defense agent · 2026

Gatehouse

Scam losses crossed an estimated $1.03 trillion worldwide in 2024. Caller ID apps label phone numbers. Banks alert after the money has moved. Nobody investigates the attack itself while it is just a message sitting on a family member's phone.

Gatehouse is that investigator. Forward the message and a team of AI agents works it like a professional fraud desk, escalating only genuine decisions to the family guardian with court-grade evidence bundles.

20
controlled design documents
30-case
offline evaluation harness
0
network calls in the rule engine
Wilson
intervals on every metric
T+0.0sFAMILYForwards the message. That is the entire user interface.
T+0.4sTRIAGE AGENTClassifies intent, extracts claims, UPI IDs, phone numbers, URLs.
T+2.1sCLAIM VERIFIERChecks registries and issuer rules from the India country pack.
T+4.8sGRAPH CORRELATORCross-household threat graph lookup, privacy-preserving aggregates only.
T+9.3sENGAGEMENT GUARDOptional guarded reply to the suspected scammer under strict rails.
T+11.7sEVIDENCE BUNDLECourt-grade packet assembled: sources, screenshots, reasoning chain.
T+12.0sGUARDIANSees one decision with evidence behind it. Everything else stays invisible.

Investigation, not classification

Agents verify claims against authoritative registries, correlate across a privacy-preserving cross-household threat graph, and can engage suspected scammers under strict guardrails. The guardian sees a decision, the evidence behind it, and nothing else.

Built on the Strands Agents SDK and deployed on Amazon Bedrock AgentCore.

Discipline before demos

Phase one runs fully offline by charter: a deterministic rule engine with zero model calls and zero network, byte-reproducible scores, and a 30-case benchmark through the same evaluation harness the learning systems must beat.

Every log record passes a mandatory scrubber, and CI seeds canary strings to prove personal data never reaches observability sinks. Metrics publish Wilson 95% intervals because small-sample honesty matters more than big-sample vanity.

Twenty controlled design documents cover vision, architecture, security, and SLOs before agent integration lands in phase two.

Next caseSignalGate