All work
AI Platform Telecom
Sentra V2: AI-Enabled Push-to-Talk
A PTT platform whose channels include an on-premise AI agent that hears, sees, and answers, with no internet connection.
- Period
- Jun 2023 – Mar 2025
- Company
- Nybsys
- Scope
- MWC 2025 demos with Intel & Lanner
Overview
Sentra is a mission-critical push-to-talk platform. In V2 I helped build the core server: users, groups, channels, and access control. Later I did much of the work that put an AI agent inside PTT channels. The agent runs speech-to-text, an LLM, and text-to-speech on local GPUs with no internet dependency. It answers from uploaded documents and photos, and can look up network faults. It was demonstrated with Intel and Lanner at MWC 2025.
Architecture
Local GPU · offline · No internet dependency
01 Client
- PTT voice
- Photo capture / upload
02 AI gateway
- Speech-to-text
- LLM / multimodal
- Text-to-speech
03 Knowledge
- Document retrieval
- Conversation memory
- Network fault lookup
What I did
01 On-prem AI agent
- Ran the full voice pipeline on a local GPU with no internet dependency: speech-to-text, the LLM, and text-to-speech.
- Built retrieval-augmented generation over documents users upload, and improved retrieval quality.
- Added multimodal questions: users can attach a photo to a query, and the system uses a larger model when a GPU is available.
- Gave the agent conversational memory, and let users interrupt a response.
- Connected the agent to the network management system so field users can ask about network faults by voice.
- Evaluated model performance and improved the quality of synthesized speech.
02 Demos & hardware
- Delivered the Intel and Lanner MWC 2025 demos on edge GPU hardware.
03 Core PTT server & platform
- Built user, group, and channel management with access control.
- Built licensing APIs and kept the web dashboard and the mobile app in sync.
- Set up Kubernetes for Sentra and made installation reliable.
- Improved real-time behaviour across the client apps.