AI Engineer

Job Title : AI Engineer

Location: Gurgaon, Haryana (On-site)

Experience: 1-3 years

What you'll build

• LLM and agentic systems that reason over operational data — DCS/historian tags, time-series, sensor streams, and ISA-95-structured objects — not only PDFs.
• Retrieval and prediction over messy industrial reality: engineering documents, P&IDs, incident reports, schematics, plus tag-level plant data.
• AI features where a wrong answer is a safety and liability event, not a UX bug. Correctness, abstention, and human-in-the-loop are the job.

Responsibilities

• Design, build, and iterate LLM-powered workflows — retrieval, routing, tool use, function calling, multi-step agents.
• Implement agentic apps that plan, call tools/APIs, and maintain state across tasks.
• Build RAG pipelines end-to-end: ingestion, chunking, embeddings, indexing, and latency-optimized retrieval — including layout-aware parsing of drawings and reports.
• Own prompt engineering and evaluation: golden datasets built with domain experts, A/B tests, guardrails, and metrics across latency, cost, quality, and safety.
• Productionize with observability (traces, tokens, failures), cost controls, and fallbacks (LangSmith / Langfuse / Arize or equivalent).
• Ship backend services and APIs (Python / FastAPI) integrating with data stores, vector DBs, and time-series sources.
• Handle deployment reality for industrial clients: VPC, on-prem, or air-gapped environments and self-hosted open models where data can't leave the plant.
• Collaborate with PM/Design and SMEs to translate requirements into reliable, safe, user-facing features.

Must have

• Hands-on with LLMs (OpenAI, Claude, Llama, etc.) and orchestration frameworks (LangChain, LlamaIndex, or custom).
• Strong Python; RESTful services; clean, tested code.
• RAG in production — vector databases (Pinecone, Weaviate, FAISS, Qdrant) and embeddings.
• Solid grasp of agent patterns: tool calling, planning/execution, memory, workflow engines.
• Prompt design, safety/guardrails, and evaluation frameworks — with a real point of view on hallucination control and confidence/abstention.
• Cloud and deployment basics (AWS/GCP/Azure), Docker, Git, CI/CD.
• Strong debugging mindset and a bias to ship.

Strong plus

• MCP — we lean on it heavily; real familiarity is a signal we're actively looking for.
• Self-hosted / on-prem inference: vLLM, quantized models, private deployment.
• Time-series, anomaly detection, or classical ML — the predictive side of industrial AI, not just LLM apps.
• Document AI and vision: OCR, layout parsing, multimodal models for drawings and reports.
• Model adaptation for cost/quality: fine-tuning, LoRA, distillation, small-model routing.
• Exposure to manufacturing, process safety, or industrial data.

Why this role

You'll build AI for a domain almost nobody has cracked — Indian process industries, at the frontier, with real clients and real plant data. The problems are hard because they're safety-critical and the data is raw. That's the point.