#9-AI Voice Stack Engineer-Dograh/Pipecat(STT, TTS, LLM)Build & Own Our Self-Hosted Voice AI

Job Type & Collaboration Type: Long-term, ongoing monthly contract (multi-year vision). Schedule: Must be available for core collaboration hours daily within 6:00 AM – 4:00 PM EST (approximately 2–4 hours per day). The Core Ask We need an engineer to build out our STT, TTS, and LLM stack directly on our own hardware (a single machine running an NVIDIA RTX 5090) inside Dograh, the open-source, self-hostable Vapi/Retell alternative (github.com/dograh-hq/dograh). This is not a "connect some API keys" job. Dograh's real-time voice runtime is built on Pipecat (the open-source real-time voice AI framework — Dograh's backend depends on it directly, via tuner-pipecat-sdk). Deep, hands-on Pipecat knowledge is a hard requirement, not a nice-to-have. You need to genuinely understand how Pipecat pipelines, frame processors, transports, and pipeline tasks work under the hood — because you'll be extending and optimizing at that layer, not just configuring a UI. What We Already Have (You Are Not Starting from Zero) We have already built and are running a real-time AI voice pipeline. Currently running: Dograh Core — the base framework is installed and operational. STT & TTS — speech-to-text and text-to-speech models integrated at a basic level. Pinecone — vector database connected to our Fusion PBX for memory/context. Custom Portal — a basic Vapi-style dashboard for managing voice agents. Hardware — single machine, NVIDIA RTX 5090, dedicated to this stack. Your Mission Take our existing Dograh instance and: Build out our self-hosted STT, TTS, and LLM stack on our own hardware — this means selecting, deploying, and tuning the actual models running on our RTX 5090 (not just wiring up third-party cloud APIs), so our voice pipeline is fast, private, and cost-controlled. Go deep on the Dograh + Pipecat architecture — understand exactly how Dograh's pipelines are constructed on top of Pipecat, where the STT/LLM/TTS frame processors plug in, and how to extend or replace them cleanly as we add features. Complete the migration to fully replicate Vapi's feature set (see modules below), so our self-hosted platform is a robust, production-grade alternative to Vapi. Required Features (We Need to Build / Complete) Module Details Assistants Create, edit, delete voice assistants (Dograh native). Squads Group multiple assistants/numbers for load balancing/IVR. Tools Function calling and custom integrations via Dograh. Phone Numbers Provision and manage inbound/outbound numbers (SIP trunking). Logs Call recordings (listen back), transcripts of assistant + user dialogue. Voice Library Manage TTS presets and custom voices. Metrics Total call minutes, number of calls, costs, etc. Integrations Configurable list of third-party integrations (CRM, etc.). API Keys Generate and manage API keys for programmatic access. If you already know Vapi's UI/API, Dograh's architecture, and Pipecat's pipeline model, you already know the target. Technical Scope Current Stack (Already Running) Real-time AI voice pipeline (STT → LLM → TTS), Pipecat-based, on RTX 5090. Dograh as the core automation framework. Pinecone for memory/context. Fusion PBX (SIP telephony). Custom portal (Vapi-style dashboard). Your Responsibilities (Sole Accountability for the AI Stack) Build & Deploy the STT/TTS/LLM Stack on Our Hardware: Stand up, configure, and continuously improve the actual speech-to-text, text-to-speech, and LLM models running locally on our RTX 5090 — this is squarely your build, not an integration afterthought. Dograh & Pipecat Depth: Work directly inside Dograh's Pipecat-based pipeline layer — frame processors, transports, pipeline tasks — to add the missing modules (Assistants, Squads, Tools, Logs, Voice Library, Metrics, Integrations, API Keys) and to extend voice-pipeline behavior as needed. Low-Latency Optimization: Use TensorRT, vLLM, or similar to optimize models for ultra-low latency on our hardware. System Integration: Ensure seamless interaction between Dograh (Pipecat pipelines), Pinecone, Fusion PBX, and the dashboard. Self-Hosted Stability: Harden the system for production use (logging, alerting, backup). Long-Term Scalability: Scale the system to multi-GPU or multi-node as we grow, and maintain uptime. Team Collaboration: Work alongside 4 other engineers who handle other parts of the platform (frontend, telephony infrastructure, devops, etc.). You hold sole decision-making authority and accountability over the Dograha instance and all AI models (STT, TTS, LLM). Required Technical Expertise (Must Demonstrate) Dograh & Pipecat (Critical, Non-Negotiable): Real hands-on experience with Pipecat's pipeline/frame-processor architecture, and ideally direct experience with Dograh itself or building on top of it. If you haven't worked in a Pipecat-based pipeline before, please don't apply — this is the core of the role. AI/ML Deployment: Proficient with PyTorch, TensorRT, vLLM, or llama.cpp. Experience fine-tuning ASR/LLM models is a plus. Real-Time Voice Pipelines: Hands-on experience building and deploying STT (e.g., Whisper, Faster-Whisper) and TTS (e.g., Coqui, XTTS, or cloud alternatives) — including running these models yourself on GPU hardware, not just calling hosted endpoints. Vapi Familiarity: Understanding of Vapi's feature set and UI/API as the functional target we're replicating. Telephony Engineering: SIP protocols, integration with PBX systems (we use Fusion PBX). Familiarity with PJSIP or Asterisk is a big plus. Backend & Dashboard: Ability to build/extend a web portal (we already have a base — you'll work on the API and frontend as needed). Infrastructure: Linux, Docker, basic GPU optimization. A security-minded approach is required. Compensation & Partnership Model Monthly Retainer: $300 USD to start, with a defined path for substantial increases as system complexity and your responsibilities grow. Profit Sharing: You will receive a share of the 50% net monthly profit pool, distributed across the 5-person engineering team. As the engineer solely responsible for the Dograh instance, STT, TTS, and LLM stack, your share of the pool will reflect the critical nature of this responsibility. This is a long-term, equity-like incentive directly tied to the value you create. Note: All code, models, configurations, and intellectual property developed during this engagement remain the sole property of the company. Why Join This Project? You are not starting from scratch — the core pipeline and Dograh instance are already installed and running. Direct profit sharing — a share of the 50% net monthly profit pool, distributed across the 5-person team, weighted for your central role. High ownership (of responsibility), high impact — you own the outcome and performance of the platform's brain: the STT/TTS/LLM stack running on our own hardware. Long-term, founding-level role — shape the architecture and grow with the company. Work on cutting-edge AI + telephony, deep in the Pipecat/Dograh stack, on real-time voice agents built on hardware you control. How to Apply Send a brief message covering: Your experience with Dograh and/or Pipecat specifically — what you've built, extended, or deployed with them (this is the most important part of your application). Your experience with real-time STT/TTS pipelines and low-latency LLM serving, including any work deploying these models yourself on GPU hardware. Any work you've done with SIP / PBX integration. Your familiarity with Vapi or similar voice agent platforms. Your availability during core hours (6 AM – 4 PM EST). Please note: You will be joining a team of 5 engineers total. While the 50% profit share is split across the team, you will have sole accountability and responsibility over the Dograh instance, the AI models (STT/TTS/LLM), and voice management — built and running on our own hardware. All intellectual property developed remains company property. This is a high-impact, high-accountability role within a collaborative environment.