We are looking for a builder who treats Large Language Models (LLMs) not as magic, but as stochastic components that require engineering rigor. As our Applied AI Engineer, you will build the "brain" of our voice agents. Your challenge is to tame the unpredictability of LLMs to create agents that can book meetings, handle objections, and update CRMs reliably, all within the constraints of a real-time phone conversation. You will move beyond simple "prompt engineering" to build sophisticated systems that use tools, manage state, and recover from errors.

Key Responsibilities:

• Building the Agentic Brain:
  Design and implement the conversation logic using Python and LLMs (OpenAI GPT-4o, Anthropic Claude 3.5). Implement sophisticated "Tool Calling" capabilities (e.g., checking calendar availability live during a call) and ensure the agent knows when to use a tool versus when to speak. Define JSON schemas for tools and handle the "hallucinated arguments" that models sometimes generate.
  
  
• Latency Optimization:
  Optimize the pipeline between the Transcriber (ASR), the Brain (LLM), and the Synthesizer (TTS) to achieve sub-second response times. Implement Streaming APIs so the agent speaks the moment it has a thought, using techniques like "speculative decoding" or "filler tokens" to mask latency.

• Voice Stack Integration:
  Work directly with Retell AI and Twilio APIs to configure the voice layer. Handle edge cases peculiar to voice: interruptions (barge-in), background noise, and ambiguous silence. Tune "Turn-Taking" models to ensure the agent does not interrupt the user too aggressively or wait too long to respond
  
  
• Evaluation & Guardrails:
  Build the "Eval Harness" by creating automated test suites that grade agent performance (e.g., adherence to script, hallucinations, data collection) using frameworks like DeepEval or custom Python evaluators. Implement guardrails (e.g., NeMo Guardrails or custom logic) to prevent the agent from discussing sensitive or off-topic subjects.
  
  
• Tool Registry Construction:
  Build the integration layer that allows agents to act in the real world—connecting them to Google Calendar, HubSpot, and Salesforce APIs. Handle OAuth flows, API rate limits, and other integration constraints.

Technical Skills & Requirements:

Must-Have (The MVP Stack):

• 3+ years of Python experience, with significant recent focus on LLMs, RAG, and Agentic Workflows.
• Modern AI Stack: Hands-on experience with LangChain (specifically LangGraph for stateful agents) or raw OpenAI API integration. Understand trade-offs between frameworks and raw orchestration code.
• API Integration: Strong experience integrating 3rd-party APIs (CRMs, Calendars). Skilled at handling rate limits, retries, and authentication (OAuth) within agent workflows.
• Python Proficiency: Write clean, typed (mypy), and tested Python code. Familiarity with FastAPI for exposing agents to the web.

Strategic Future-Proofing Skills:

• Telephony Awareness: Experience with Twilio Programmable Voice, Vapi, or Retell AI. Ability to debug call flows using Twilio debug logs is a strong advantage.
• Durable Execution: Interest in learning Temporal.io. Understand that an agent is a long-running process, not just a single HTTP request, and be eager to orchestrate it reliably.

Candidate Profile:

• Pragmatic tinkerer: Likely built Discord bots, voice assistants, or automation scripts. Understand that "95% accuracy" in a demo is easy, but "99.9% reliability" in production is the real engineering challenge.

Other Details:

• Job Timings: 9 working hours between 12 pm and 12 am.
• Office location: Off to Shahrah-e-Faisal, PECHS, Karachi