Senior AI Engineer

Mirai Cyber; Pakistan; On-site; PKR 3,000,000 - 5,000,000; Full time

Job summary

MIRAI AI in Pakistan is seeking a Senior AI Engineer to design and ship end-to-end AI systems for threat intelligence, including multi-agent orchestration, RAG pipelines, and self-hosted model inference.

You will own prompt and context design, tool calling, evaluation, and production reliability with a focus on security, grounding, and scalable deployment across tech stack.

Qualifications

• Expert Python with strong async experience using FastAPI and Pydantic.
• Agentic systems in production: multi-agent orchestration, tool calling, memory, and failure recovery.
• Advanced RAG architectures: hybrid search, chunking, reranking, citation, and hallucination reduction.
• Vector databases experience: pgvector, Qdrant, Milvus, or Weaviate.
• Knowledge graphs & graph-based reasoning: Neo4j for retrieval and threat-actor correlation.
• Prompt & context engineering with structured outputs: versioned, schema-constrained LLM outputs.
• LLM inference & serving: self-hosted deployment (Ollama, vLLM) on GPU with quantization and batching.
• AI evaluation discipline: golden datasets, precision/recall/F1, CI-integrated RAG evaluation.
• AI security: defend against prompt injection and data exfiltration; manage untrusted content.
• Database fundamentals: PostgreSQL (with pgvector) and Redis.
• Production delivery: containerised services with Docker and CI/CD.

Responsibilities

• Shape, build, and operate agentic and RAG pipelines as production services.
• Make evaluation a CI-gated practice with golden datasets and regression tests.
• Drive down hallucination and improve grounding, citation validity, and retrieval accuracy.
• Ship LLM services with structured outputs and reliable tool calling.
• Harden the pipeline against prompt injection and treat data as untrusted input.
• Instrument tracing, logging, per-tenant metering, and cost/latency dashboards.
• Collaborate with CTI, product, and platform teammates to turn intelligence requirements into capabilities.

Skills

Python

Async programming

FastAPI

Pydantic

Agentic systems

LangGraph

LlamaIndex

Tool calling

RAG architectures

Vector databases

Knowledge graphs

Schema-constrained prompts

LLM inference

AI evaluation

AI security

PostgreSQL with pgvector

Redis

Docker

CI/CD

Tools

LangGraph

LlamaIndex

Autogen

FastAPI

Pydantic

Ollama

vLLM

Docker

CI/CD

Job description

MIRAI AI is a predictive Cyber Threat Intelligence (CTI) platform that combines LLM-powered agents, retrieval augmented reasoning, and multi-tenant enterprise SaaS. We're hiring a Senior AI Engineer to design and ship the systems at the core of the product: multi-agent orchestration, advanced RAG pipelines grounded in threat intelligence, self-hosted and frontier-model inference, and the tooling that connects our agents to live security data. This is a hands-on, deeply technical role. You'll own AI systems end-to-end - prompt and context design, retrieval, tool calling, inference, and the evaluation that keeps all of it honest - and get them running reliably in production. We're looking for someone genuinely deep in applied LLM/agent/RAG engineering, and comfortable across the surrounding stack - including model serving and the AI-specific security a threat-intelligence pipeline demands. You don't need to be a platform/infrastructure specialist or an application-security engineer; we build those functions around you.

What You'll Build

• Agentic systems - multi-agent workflows (planning, reasoning, human-in-the-loop, memory, retry and failure recovery) that enrich IOCs, map activity to frameworks like MITRE ATT&CK, and reason over threat actors and campaigns.
• Advanced RAG - hybrid retrieval (vector + keyword) over a graded, per-tenant intelligence corpus, with reranking, strong citation, and evidence-grounding so every claim is traceable to a source.
• Production LLM services - backends serving structured, tool-calling LLM workflows with Pydantic schema validation and graceful degradation.
• Inference & model serving - a two-tier model strategy: frontier APIs for reasoning and report generation, and self-hosted 7B-14B models (Ollama, vLLM, or equivalent) for high-volume classification, extraction, and summarization - with quantization, batching, and caching for cost and latency at volume.
• Agent tooling - tools, function-calling interfaces, and MCP server integrations that let agents access CTI data sources with tightly scoped permissions.

What You’ll Do

• Shape, build, and operate agentic and RAG pipelines as production services, not prototypes.
• Make evaluation a first-class, CI-gated practice: golden datasets, RAG metrics (e.g., RAGAS), regression tests, and red-teaming of agent and retrieval behaviour - wired in so regressions fail the build.
• Drive down hallucination and improve grounding, citation validity, and retrieval accuracy against measured outcomes.
• Ship LLM services with structured outputs and reliable tool calling, treating prompt and context engineering as versioned, tested artifacts - not ad-hoc strings.
• Harden the pipeline against prompt injection, and treat all ingested intelligence as untrusted input at every layer.
• Instrument everything - tracing, structured logging, per-tenant token metering, and cost/latency dashboards – so quality regressions and drift are caught before customers see them.
• Collaborate with CTI, product, and platform teammates to turn intelligence requirements into reliable AI capabilities.

Must-Have Qualifications (Core)

These are the non-negotiables - we expect real production depth here.

• Expert Python - strong async experience, building services with FastAPI and Pydantic.
• Agentic systems in production - multi-agent orchestration, tool calling, agent memory, state/checkpointing, and failure recovery - with a modern framework such as LangGraph (or a defensible equivalent: LlamaIndex agents, Autogen, custom orchestration).
• Advanced RAG architectures - hybrid search, chunking strategies, reranking, retrieval evaluation, citation, and hallucination reduction.
• Vector databases - practical experience with pgvector (today's default), and/or Qdrant, Milvus, or Weaviate as scale requires.
• Knowledge graphs & graph-based reasoning - modelling entities and relationships (e.g., Neo4j) for retrieval and threat-actor correlation.
• Prompt & context engineering with structured outputs - reliable, schema-constrained LLM behaviour through tool calling; prompts and context handled as versioned, tested artifacts.
• LLM inference & serving - self-hosted deployment (Ollama, vLLM, or equivalent) on GPU, plus quantization, batching, caching, and latency/cost optimisation at volume.
• AI evaluation discipline - golden datasets, precision/recall/F1, RAG evaluation (e.g., RAGAS), and regression testing wired into CI.
• AI security - defending against prompt injection and data exfiltration via tool use, scoping tool permissions, handling untrusted ingested content, and awareness of model supply-chain risk.
• Database fundamentals - PostgreSQL (with pgvector) and Redis.
• Production delivery - shipping containerised services with Docker and CI/CD.

Experience

• 5+ years of software engineering.
• 3+ years shipping production LLM / generative-AI applications, including hands-on RAG and agent work.
• Experience contributing to a scalable, multi-tenant SaaS product.

Strong Plus

You won't have all of these - depth in a few is what we're after.

• MCP (Model Context Protocol) - building MCP servers and custom tools.
• Hugging Face & PyTorch - inference optimisation, fine-tuning (LoRA/QLoRA), embeddings, and reranking models.
• Search infrastructure - OpenSearch / Elasticsearch.
• Observability stack - OpenTelemetry, Prometheus, Grafana, distributed tracing, AI-quality dashboards.
• CTI data substrate - working with OpenCTI, MISP, MITRE ATT&CK STIX, CVE/NVD, EPSS, CISA KEV, and SSVC.

Nice to Have (Optional)

Genuinely optional. Strength here is a bonus, not an expectation - we don't expect one person to own the platform or security functions.

• Platform / infrastructure - Kubernetes, Helm, Terraform, AWS or Azure, autoscaling, distributed and GPUcluster inference.
• Application & platform security - OAuth2/OIDC, JWT, RBAC/ABAC, secret management, multi-tenant isolation, audit logging, secure API design.
• CTI domain knowledge - threat actors and campaigns, IOC enrichment, TTPs, detection-engineering basics, SIEM/XDR integrations, dark-web and EASM exposure. (You'll pick a lot of this up on the job.)
• Frontend - React, Next.js, TypeScript, Tailwind, data visualisation for internal tooling and dashboards.