URGENT: LiveKit + Azure Voice AI Expert Needed: Latency Tuning & Prompt Optimization (3 hour task)
Publicada el 2026-07-18
Descripción de la oferta
We have a fully built, 100% functional, and Dockerized Voice AI infrastructure (Inbound Receptionist & Outbound Sales Dialer) built on the LiveKit Python SDK and Azure (STT, TTS, OpenAI GPT-5.6-LUNA). The SIP routing, databases, and core connections are completely flawless. We need a LiveKit/Voice AI expert for a quick 2 to 3-hour optimization task. The agents are successfully taking and making calls, but the AI behavior and latency need expert tuning. If you are a LiveKit expert, this is an incredibly easy "drop-and-pick" task. The 3 Tasks Required: 1. Latency & VAD Tuning (Speed) The current LLM Time-To-First-Token (TTFT) is occasionally hitting 2.5 - 3+ seconds. TTS is already blazing fast (0.2s). Your task: Tune the LiveKit AgentSession parameters, VAD endpointing, and Azure OpenAI stream settings to bring the perceived latency down to human-like levels (< 700ms). 2. Strict Prompt Engineering (Brevity & Guardrails) Currently, the AI talks too much (sometimes rambling for 30 seconds). Your task: Rewrite the system prompts to force extreme brevity (1-2 sentences max). 3. Tool Calling & Hangup Logic Fixes Outbound Bug: The agent has a log_call_disposition tool that natively hangs up the call when finished. Right now, if a user asks "How much to ship this to Chennai?", the AI hallucinates, assumes the order is placed, fires the tool, and hangs up mid-conversation. You must fix the prompt/tool description so it never hangs up until the user explicitly confirms the order. Inbound Feature: The inbound agent currently stays on the line indefinitely. You need to implement a native, graceful hangup trigger (via tool call or LiveKit command) so the AI ends the call when the customer says "Goodbye" or "Thank you, that's all." How You Will Work (Strict Testing Rule): I will not be testing your iterations for you. You must: Spin up the provided files using your own LiveKit Cloud project and Azure API keys. Call the agents yourself to test the latency, the tool calls, and the hangup logic. Once you are 100% certain the AI behaves perfectly, responds instantly, and never hangs up prematurely, deliver the updated files. Files We Will Provide: livekit_agent.py (The Inbound Worker) outbound_agent.py (The Outbound Worker) ai_core.py (Contains the FactoryManagementTools class and Base Prompts) Requirements: Deep experience with LiveKit Python SDK (specifically the latest Agent framework, room_options, and Native Timeouts). Experience minimizing latency with Azure OpenAI & Azure TTS. Expertise in strict LLM prompt engineering for tool-calling. The files will be provided to the awarded freelancer only after verified that they are human, and an actual expert in the required field. The task is required to be completed in 1 DAY COMPULSORY . ITS A QUICK TASK FOR ANYONE WHO IS AN EXPERT IN THE FIELD. THE FREELANCER SHOULD HAVE: LIVEKIT CLOUD CREDENTIALS : TO HOST AN AGENT IN AP-SOUTH-MUMBAI REGION AZURE CREDENTIALS: TO SPIN UP AZURE SPEECH SERVICES IN CENTRAL INDIA, AND AZURE OPENAI SERVICE WITH A GPT-5.6-LUNA WITHIN MICROSOFT FOUNDRY, DEPLOYED IN GLOBAL STANDARD IN THE SOUTHEAST-ASIA REGION. THESE ARE THE MANDATORY PREREQUISITES. WITHOUT THIS, YOU'RE NOT ELIGIBLE TO BE AWARDED. Deliverables: You have to deliver the key codebase ALONG WITH: A transcript.txt, STRAIGHT FROM LIVEKIT CLOUD, SHOWING A CALL WHERE THE LATENCY REACHED SUB 700-800MS, THE ENTIRE CALL FLOW WAS HANDLED PERFECTLY. One transcript for each inbound and outbound agent. Please start your proposal with "LIVEKIT AZURE" so I know you read the requirements. I am ready to hire immediately and hand over the codebase.
Skills
Fuente original: freelancer