Job Description

AI Engineer — Agentic AI, RAG & Low-Latency Voice Systems

n

Mobiloitte Technologies (I) Pvt. Ltd. | New Delhi (on-site / hybrid) | Full-time | 5+ years

n

Compensation: No ceiling for the right candidate.

n

n

Two things decide this role: how well you handle latency, and how far past basic RAG you can build.

n

n

Mobiloitte builds production AI for enterprise clients across India, the US, UK, UAE, Singapore and South Africa. We're hiring a senior AI Engineer to own our agentic AI, retrieval and voice stack — systems that go live, carry real call volume, and stay up.

n

n

1. LATENCY IS THE JOB, NOT A DETAIL

n

n

Our voice bots handle inbound and outbound calls where a 400ms difference decides whether the conversation feels human. We need someone who treats latency as an engineering discipline:

n

n

- Owns end-to-end voice latency from user speech to first token of bot audio — and can quote the numbers they achieved

n

- Profiles and attacks each hop: STT, retrieval, inference, TTS, network

n

- Streaming responses, speculative/partial generation, response chunking to shorten time-to-first-audio

n

- Barge-in and interruption handling that actually works mid-sentence

n

- Model caching, warm pools, batching, concurrency control under real call load

n

- Knows when the fix is architectural, not a bigger GPU

n

n

If you can't tell us what your p95 latency was and what you did to bring it down, this isn't the role.

n

n

2. AGENTIC AI — BEYOND BASIC RAG

n

n

Single-shot retrieve-and-answer is table stakes. We're building systems that plan, act and self-correct:

n

n

- Agent orchestration — LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, or your own framework (we care that you've built it, not which one)

n

- Tool and function calling, API/DB actions, structured outputs

n

- Multi-step reasoning, task decomposition, planning loops with retry and fallback

n

- Multi-agent patterns — routing, supervisor/worker, handoffs

n





- Conversation and long-term memory management across sessions

n

- Advanced retrieval: agentic and query-rewriting RAG, hybrid search, reranking, graph/multi-hop retrieval, chunking strategy chosen deliberately

n

- Evaluation and guardrails — hallucination reduction, groundedness scoring, tracing and observability (LangSmith, Langfuse or similar)

n

- Cost control: token budgeting, context-window discipline, model routing

n

n

3. VOICE — INBOUND AND OUTBOUND

n

n

- Full pipeline: STT → LLM/agent/RAG → TTS

n

- Whisper / Faster-Whisper / Deepgram; Piper / Coqui / XTTS or equivalent

n

- SIP and telephony integration, WebSockets, real-time streaming

n

- Built the stack rather than wrapping Vapi, Retell or Bland end to end

n

n

4. PRIVATE / SELF-HOSTED DEPLOYMENT

n

n

Many of our clients cannot send data to public APIs. You should be able to:

n

n

- Deploy Llama, Qwen or Mistral inside a client VPC or on-prem

n

- Run vLLM, Ollama or TensorRT-LLM in production

n

- Size GPUs and VRAM, apply INT8/INT4 quantization, tune inference

n

- Handle concurrency, autoscaling, secure API exposure, auth and data isolation

n

- Explain how you'd replace an OpenAI dependency with a self-hosted architecture — this earns robust preference

n

n

ON COMPENSATION

n

n

We've deliberately not published a band. For an engineer who has genuinely built agentic systems, driven latency down in production, and run models on their own infrastructure, compensation is not the constraint — we'll match what the work is worth.

n

n

Be honest with yourself before applying: if your AI work is mainly calling third-party APIs and writing prompts, we'll find that out in the first twenty minutes. If you've built agents, fought latency, and put a model on your own GPU, we want to talk.

n

n

HOW TO APPLY

n

n

Apply here and include one thing: a link, repo, demo video or short architecture note for a system you personally built — and tell us what you did on it.

n

n

Thanks

n

Team-HR

📌 AI Engineer -- Agentic AI, RAG & Low-Latency Voice Systems | Okhla | 5+ yrs
🏢 Mobiloitte
📍 Okhla