The posting
Role overview
We are hiring AI engineers to build agentic AI systems: LLM agents that plan, call tools and complete real
business tasks end to end. You will build agents that hold live phone and web conversations —
streaming speech recognition, reasoning and tool calls, and speech synthesis in real time. You will also
own the evaluation system that proves each agent works before every release, and help move use cases
from a demo to a reliable, observable service that clients trust with their customers.
Key responsibilities
Agentic systems
• Design and build LLM agents that plan, call tools and complete multi-step business tasks across
voice, chat and back-office channels.
• Implement task and conversation state, memory and context management, retrieval (RAG) and
model routing between fast and more capable models.
• Define tool schemas and permissions, validate every tool call, and add human approval points
for high-impact actions.
• Build durable workflows with checkpoints, retries, idempotency keys, timeouts and
compensation for failed actions.
• Use multi-agent designs, such as a router with specialist agents, only where they measurably
beat a single agent.
• Integrate CRMs, ticketing systems and databases for actions such as complaint registration,
classification, routing, status updates and notifications, with warm human handoff.
Voice agents
• Build streaming pipelines that connect telephony (SIP, Twilio, Exotel) and WebRTC audio to ASR,
LLMs and TTS over WebSockets.
• Implement voice activity detection, turn detection, barge-in and cancellation so the agent stops
talking the moment a caller speaks.
• Set and hold latency budgets per hop (ASR, LLM, TTS, network), with a target of under [1 s] from
end of caller speech to first agent audio.
• Adapt agent behaviour for speech: short turns, confirmations before actions, and recovery from
mishearing.
• Support multilingual and code-switched calls, including language detection and matching TTS
voices.
Evaluation
• Build and maintain golden test sets that define the user goal, persona, expected tool calls and
pass criteria for each scenario.
• Run trajectory evals that check each step's tool choice, arguments and order, not just the final
answer.
• Run simulated-caller evals for voice: an LLM plays the caller, TTS voices it with accents, noise
and interruptions, and the agent is tested end to end over real audio.
• Score task success, tool-call correctness, policy compliance, hallucination, word error rate and
latency, using rule checks and LLM judges calibrated against human ratings.
• Gate releases in CI on eval results, block regressions, and turn every confirmed production
failure into a new test case.
Backend and operations
• Write async Python services (FastAPI, asyncio) with structured logging, tracing and metrics for
every agent run and conversation.
• Load-test concurrent sessions, find bottlenecks and keep p95 latency and per-task model cost
within budget.
• Protect sensitive data: authentication and authorisation, audit trails for agent actions, PII
redaction and tenant isolation.
• Ship with Docker and CI using staged releases and rollback, and troubleshoot live agents in
production.
• Work with product and delivery teams to turn client workflows into working agent integrations.
Required qualifications and skills
• Strong Python and backend skills: APIs, asyncio, WebSockets, structured error handling and
automated tests.
• Hands-on experience shipping LLM agents beyond a prompt demo: tool calling, state
management, RAG and failure handling in a real workflow.
• Sound judgement on agent design: when a fixed workflow beats an autonomous agent, single
versus multi-agent, and how to keep agents safe with permissions and guardrails.
• Working knowledge of a real-time voice stack: streaming ASR and TTS, VAD, turn-taking and the
trade-offs between latency and accuracy.
• Experience designing evaluations for LLM agents: test sets, trajectory and outcome metrics,
LLM-as-judge and regression tracking — and knowing when a metric is misleading.
• Ability to debug across services using logs, traces and recorded sessions to find where an agent
failed or slowed down.
• Experience with SQL or NoSQL databases, Git, Docker and CI pipelines.
• Bachelor’s degree in a technical field or equivalent practical experience, with clear examples of
delivered work.
Preferred experience
• Agent frameworks and orchestration tools such as LangGraph, OpenAI Agents SDK, Temporal or
MCP tool servers.
• Production voice agents on telephony (SIP, Twilio, Exotel) or WebRTC, with frameworks such as
LiveKit Agents or Pipecat.
• Speech providers such as Deepgram, Whisper, Google STT, ElevenLabs, Azure TTS or Cartesia,
and measuring their WER and latency on your own data.
• Indian-language and code-switched speech (Hindi, Telugu, Tamil and English).
• Queues, caching and event-driven systems (Redis, Kafka, RabbitMQ), and load-testing tools such
as Locust or k6.
• Access controls, PII redaction and on-premises or restricted-network deployments.
• Other modalities, such as vision or document understanding (images, screenshots, scanned
forms), and combining them with voice or text in one agent.
Success in the role
Success means AI agents — voice, chat or back-office — that complete real business tasks reliably, with
evidence that they work and a codebase the team can build on. In your first six months, that looks like:
• Business outcomes: agents measurably resolve tasks without human help, cut handling time
and raise customer satisfaction for client use cases.
• Reliability: workflows recover from timeouts, tool failures and bad input; write actions are
never duplicated; every incident is traced to a cause and fixed.
• Evaluation culture: every release is gated by evals, metrics track what users actually experience,
and production failures flow back into the test sets.
• Performance and cost: latency and per-task model cost stay within agreed budgets as traffic
grows.
• Delivery: client requirements become working integrations on schedule, with clear handover to
product and delivery teams.
• Engineering quality: features are tested, observable, documented and easy for others to
maintain; you take an active part in code review and shared tooling.
