Talklyatics - Talk to Your Data - Get Insights Instantly

Abstract

Talklyatics is a security-first Multi-Agent System (MAS) designed to enable safe, reliable, and auditable execution of Large Language Model (LLM)-generated analytical code for conversational data analytics. Existing LLM-based analytics platforms suffer from execution security gaps, hallucinated code outputs, speculative self-correction, and insufficient auditability rendering them unfit for enterprise deployment. Talklyatics addresses these limitations through three core architectural innovations: (1) Dynamic Policy Enforcement (DPE), an Abstract Syntax Tree (AST)-based pre-execution static analysis layer that blocks all prohibited operations before sandbox invocation; (2) In-Execution Self-Correction, a deterministic, state-aware feedback mechanism that captures intermediate runtime variable snapshots to guide targeted, context-preserving code regeneration aligned with Quality-aware Cost Optimization (QC-Opt) principles; and (3) Enhanced Container Isolation (ECI), a hardened Docker-based sandbox enforcing strict cgroup resource limits, read-only file system whitelisting, seccomp syscall filtering, and zero external network access. Empirical evaluation demonstrates a 0% DPE policy bypass rate, a 0% container breach rate, a 65% autonomous runtime error resolution rate within two correction iterations, and a 20-30% reduction in LLM API token consumption relative to unconditional session-reset baselines. The system further supports multilingual interaction (English, Hindi, Marathi) and fully automated visualization generation, establishing a replicable architecture for production-grade conversational analytics.