AI / ML2026

VoiceDesk

A local-first voice agent with hybrid retrieval, working and long-term memory, guardrails, MCP-shaped tools, and a draft-confirm-execute workflow that prevents unapproved state changes.

PythonAgentic WorkflowsBM25 RAGMCPASR/TTSGuardrails

VoiceDesk is a provider-agnostic Tanglish service-desk voice agent designed around a simple rule: natural conversation should never bypass explicit control over real-world actions.

Provider-neutral by design

Typed ports separate the core workflow from model, speech, retrieval, and tool providers. The agent can run in a zero-key local mode or connect to hosted services without rewriting its business logic.

Retrieval and memory

BM25 retrieval grounds service-desk answers in the available knowledge base. Working memory maintains the current conversation, while long-term memory preserves useful context without making every past interaction part of every prompt.

Confirmation-gated tools

State-changing requests follow a draft → confirm → execute lifecycle. The agent first explains the intended action and its parameters, then waits for explicit approval before invoking an MCP-shaped tool. Guardrails prevent the conversational layer from silently skipping that boundary.

Tanglish voice workflow

ASR and TTS adapters support a natural spoken workflow for Tamil-English users while keeping the orchestration layer independent of any single speech provider.

Designed & built withby Dhevenddra

Built with Next.js, Tailwind CSS & Framer Motion