AI Engineer - Voice & Agentic Systems

Job Title: AI Engineer (Voice & Agentic Systems) Location: Noida Working Mode: Remote Type: Full-time Experience: 2+ Years (shipping LLM agents) About the Company: weya AI builds AI automations

• Voice, video and WhatsApp agents that banks, lenders and insurers put in front of their customers. Our agents handle collections, onboarding, renewals, KYC follow-ups and support calls at scale, across India, the Middle East & Africa, Australia and the US. We're a small team of 11-15 people. Every engineer ships to production, talks to customers, and owns real outcomes. If an agent mishandles a loan call, it costs a client money and trust, so we build for reliability first.

About the Role

You'll design, build and run the agents that talk to our clients' customers: LLM orchestration with Llama. Index/Lang. Chain, FastAPI services behind them, and real-time voice pipelines wired into telephony providers. You own an agent end to end, from the first prompt to the 10,000th concurrent call. What You'll Do? Build agentic workflows with Llama.

Index and Lang. Chain: tool calling, RAG over policy and loan documents, multi-step reasoning, and guardrails for regulated conversations. Design and run FastAPI services that serve agents in real time: async endpoints, Web. Sockets, webhooks, background tasks and clean API contracts. Build low-latency voice agents end to end: STT, LLM, TTS, voice activity detection, barge-in/interruption handling, and turn-taking.

Integrate and maintain telephony providers (e.g. Twilio, Plivo, Exotel, Vonage, SIP trunks): inbound/outbound calls, call transfer, recordings, DTMF, and media streaming. Engineer for failure: provider outages, dropped calls, LLM timeouts, partial transcripts, retries and idempotency, so one bad dependency never takes down a campaign.

Write tests that catch problems before clients do: unit and integration tests, simulated calls, load tests, and regression suites for agent behaviour. Instrument everything: latency per turn, call success rates, cost per call, and alerts that fire before a client notices. Work directly with FDEs and clients to turn real call transcripts into better agents.

Requirements

What We're Looking For? (Non-Negotiable) Agentic frameworks: shipped production agents with Llama. Index and/or Lang. Chain (or Lang. Graph), including tool use, memory and RAG. Not just notebooks. FastAPI: extensive experience building async Python services with FastAPI, Pydantic, Web. Sockets and background workers, under real traffic.

Voice agents: built at least one voice agent that handled real customer calls, with hands-on work on STT/TTS providers (e.g. Deepgram, Eleven. Labs, Azure, Google) and latency tuning. Telephony: integrated two or more telephony providers or SIP setups, and can explain how audio flows from a phone network to your code and back.

Production mindset: you think about concurrency, rate limits, timeouts, retries and edge cases before writing the happy path, and you write tests for them. Debugging under pressure: comfortable reading logs, call recordings and traces to find why a call failed. Clear written English; able to explain a technical trade-off to a client.

Good to have (Bonus Points): Real-time voice frameworks such as Live. Kit Agents or Vapi, or speech-to-speech models. Go for high-throughput services alongside Python. Multilingual voice work, especially Hindi, Hinglish, Arabic or other Indic languages. Experience in BFSI or other regulated domains: consent, call recording rules, PII handling, audit trails.

LLM evaluation: automated scoring of agent conversations, prompt regression testing. WhatsApp Business API or video agent experience. Docker, Kubernetes, AWS/GCP, Redis, Postgres, and queues like Kafka or RabbitMQ.

Benefits

Get Real Experience on "What Success looks like?" To apply: send your CV at [email protected], GitHub or a demo of a voice agent you built, and one paragraph on the hardest telephony problem you've solved. Open question: application email or link.