AI Architect

Mindsprint Pte. Ltd.; Chennai; On-site; INR 2,400,000 - 4,800,000; Full time

Job summary

Mindsprint Pte. Ltd. is seeking an AI Architect to own the end-to-end technical architecture of an enterprise-grade agentic AI platform. You will define patterns, guardrails and reference implementations enabling AI product teams to build safely at scale.

You will lead architecture reviews, design ADRs, mentor engineers and evaluate new frameworks, models and vendors to optimize ROI and deployment speed across multiple use cases.

Qualifications

• 10–14 years in software engineering/architecture, with 4+ years in ML/AI and 2+ years shipping LLM-based or agentic systems to production.
• Deep hands-on knowledge of LLM application design: prompting, function/tool calling, structured outputs, context management, RAG.
• Experience with agent frameworks and major LLM providers (Anthropic, OpenAI, Google, Azure OpenAI, AWS Bedrock).
• Strong distributed-systems background: microservices, event-driven design, APIs, queues, async workflows.

Responsibilities

• Define reference architecture for single- and multi-agent systems, including orchestration, planning, tool use, memory and human-in-the-loop checkpoints.
• Choose agent patterns for various use cases and design the tool/integration layer (MCP servers, contracts, enterprise APIs).
• Architect retrieval and memory (RAG pipelines, vector/graph stores).
• Set model strategy and cost/latency trade-offs; define LLMOps/AgentOps stack and evaluation frameworks.

Skills

LLM architecture

Agentic AI

Python

Distributed systems

Context management

Education

B.E./B.Tech./M.Tech/MS in CS/AI

Tools

LangGraph

LangChain

LlamaIndex

CrewAI

Semantic Kernel

OpenAI Agents SDK

Job description

Experience: 10–14 years total; 4+ years in ML/AI, including 2+ years building LLM and agentic systems in production

Function: Core AI Platform / Agentic AI CoE

Reports to: Head of AI / Chief Architect

Employment: Full-time

Role summary

The AI Architect owns the end-to-end technical architecture of our agentic AI platform: how agents reason, plan, call tools, remember, and stay safe at enterprise scale. You set the reference architecture, patterns and guardrails that every AI product team builds on, and you are the final technical authority on agent design choices.

Key responsibilities

Architecture and design

• Define the reference architecture for single- and multi-agent systems: orchestration, planning, tool use, memory, state, and human-in-the-loop checkpoints.
• Choose and justify agent patterns (ReAct, plan-and-execute, supervisor/worker, graph-based workflows) for each use case.
• Design the tool and integration layer, including Model Context Protocol (MCP) servers, function-calling contracts and enterprise API access.
• Architect retrieval and memory: RAG pipelines, hybrid search, vector and graph stores, short- and long-term agent memory.
• Set the model strategy: frontier API models vs. open-weight models, routing, fallbacks, fine-tuning vs. prompting, and cost/latency trade-offs.

Platform, scale and reliability

• Design for production: latency budgets, concurrency, caching, token cost controls, and graceful degradation.
• Define the LLMOps/AgentOps stack: tracing, evaluation, prompt and model versioning, observability and rollback.
• Establish evaluation frameworks covering task success, tool-call accuracy, hallucination rate, safety, and regression testing.

Governance, security and Responsible AI

• Build guardrails against prompt injection, data leakage, unsafe tool actions and over-permissioned agents.
• Define identity, authorization and audit for agents acting on behalf of users.
• Align designs with data privacy and AI regulation (e.g. DPDP Act, GDPR, EU AI Act where relevant) and internal Responsible AI policy.

Leadership and influence

• Run architecture reviews and write design docs, ADRs and reference implementations.
• Translate business problems into agentic solutions with clear ROI with product and business stakeholders.
• Evaluate emerging frameworks, models and vendors, and decide build vs. buy.
• Mentor senior engineers and grow architecture capability across teams.

Must-have qualifications

• 10–14 years in software engineering/architecture, with 4+ years in ML/AI and 2+ years shipping LLM-based or agentic systems to production.
• Deep hands-on knowledge of LLM application design: prompting, function/tool calling, structured outputs, context management, RAG.
• Working experience with agent frameworks such as LangGraph, LangChain, LlamaIndex, CrewAI, AutoGen, Semantic Kernel, OpenAI Agents SDK or Claude Agent SDK.
• Experience with major LLM providers (Anthropic, OpenAI, Google, Azure OpenAI, AWS Bedrock) and open-weight models (Llama, Mistral, Qwen).
• Strong distributed-systems background: microservices, event-driven design, APIs, queues, async workflows.
• Vector databases and search: Pinecone, Weaviate, Milvus, pgvector, Elasticsearch/OpenSearch.
• Cloud architecture on at least one of AWS, Azure or GCP, with Kubernetes and containerisation.
• Strong Python, working knowledge of Java, TypeScript or Go.
• Proven record of owning architecture decisions with measurable business outcomes.

Good to have

• MCP server design and agent-to-agent (A2A) protocols.
• Fine-tuning and alignment techniques: LoRA/QLoRA, SFT, DPO; model serving with vLLM, TGI or Triton.
• Knowledge graphs and GraphRAG.
• Observability tools such as LangSmith, Langfuse, Arize Phoenix, Weights & Biases.
• Cloud architect certification (AWS / Azure / GCP); publications, patents or open-source contributions in AI.
• Domain exposure to BFSI, healthcare, retail, manufacturing or supply chain.

Education

B.E./B.Tech/M.Tech/MS in Computer Science, AI, Data Science or a related field. PhD a plus, not required.

Success in the first 12 months

• A published agentic reference architecture adopted by all AI product teams.
• 3+ agentic use cases moved from pilot to production with defined SLAs and evaluation gates.
• A guardrail and evaluation framework in place, with measurable drops in hallucination and unsafe-action rates.
• Clear reduction in LLM cost per transaction through routing, caching and model selection.