CLOSE
aigocy@gmail.com

CUP

TWITTER (X) DRIBBBLE LINKEDIN

Configuration

COLORS
CUSTOM CURSOR
Skip to main content

espialtech

Menu
Start a Project

VoiceDesk

Home
/
Blog
espial
July 27, 2026
No Comments
VoiceDesk
Private Voice AI

Everything a voice assistant should do —
and nothing leaves your network.

Project Details

Voice assistants need low-latency speech-to-text, reliable intent detection and safe action execution. Most deployments get there by outsourcing inference to a third-party API, which is exactly what a regulated client cannot do. VoiceDesk runs entirely on a self-managed VM inside the customer boundary with controlled network egress. Whisper handles speech-to-text with streaming output, so the transcript appears incrementally and the interaction feels conversational rather than laggy.

Intent detection routes each utterance to a structured action candidate, and a policy layer validates permissions before any app operation is invoked — intent, then policy check, then execution, never in a different order. Private retrieval over a local vector database answers app-specific questions and surfaces navigation steps. Multilingual support routes intent in a language-aware way rather than translating first and losing the meaning.

DELIVERABLES
Self-hosted STT Intent routing Permission-gated actions Multilingual support
INDUSTRY
Financial Services / Regulated Enterprise

Project Research

The design constraint was absolute: zero third-party inference dependency.

That ruled out the easy path and forced open-weight models (Whisper for STT, Llama/Mistral for reasoning) running under Docker on the customer VM. The engineering challenge became latency — keeping a self-hosted end-to-end voice loop under two seconds on hardware we did not choose.

Project Results

VoiceDesk delivers an end-to-end voice loop in under two seconds, with 85% of tasks completable via voice suggestions, multilingual support, and — the number that mattered most to the client — zero third-party inference dependency. Every inference endpoint stays inside the customer VM boundary.

LedgerLens
Finance Document AI
View Project

LedgerLens
Finance Document AI

DESCRIPTION
Reads invoices, POs, receipts and bank statements in any format, matches them automatically, flags anomalies and posts clean entries to the ERP. 94% touchless, $1.1M in leakage prevented.
DELIVERABLES
Layout-aware extraction, 3-way matching, Anomaly detection, ERP posting
INDUSTRY
Finance / Back Office