Overview

We are looking for a Lead Gen AI & Agentic AI Engineer to own the design and delivery of LLM-powered and autonomous agent systems for enterprise clients.

Responsibilities

• Own end-to-end design of Gen AI solutions: retrieval architecture, model selection, prompt strategy, and integration approach
• Design agentic systems: tool contracts, orchestration logic, state management, and multi-agent decomposition where it is justified
• Define the boundaries of autonomy: what an agent may access, what it may act on, where a human must approve, and how it escalates
• Build guardrails against prompt injection, sensitive data exposure, and unsafe or incorrect output
• Design failure behaviour explicitly: retries, timeouts, circuit breakers, cost ceilings, and rollback paths
• Establish evaluation frameworks for LLM and agent output, and use them to drive iteration rather than relying on impressions
• Treat token and inference cost as a first-class design constraint
• Work directly with client stakeholders to shape requirements and explain technical trade-offs
• Support pre-sales and solutioning by estimating effort, assessing feasibility, and identifying risk early
• Set technical direction on engagements and review the team's work against it
• Hand over production-ready systems with the necessary instrumentation and documentation
• Mentor associate engineers through code review, pairing, and design discussion
• Assist in executing Binariq's internal training on Gen AI and agentic engineering
• Own the technical onboarding path for engineers joining the practice, so new joiners reach useful output quickly
• Set and enforce engineering standards across the team, covering evaluation, guardrails, and safe agent design
• Build reusable patterns, reference implementations, and internal documentation the team can work from

Qualifications

• Strong Python, including asynchronous programming, packaging, and testing discipline
• FastAPI or equivalent, with sound API and microservices design
• SQL and NoSQL, with judgement about which fits a given problem
• System design: how the AI components sit within a wider client architecture
• Production experience with multiple LLM providers, and informed opinions about their trade-offs
• Advanced RAG: chunking strategy, embedding model selection, hybrid and semantic search, re-ranking, query rewriting, and diagnosing retrieval failure
• Vector databases: practical experience, including indexing strategy and retrieval performance at scale
• Prompt and context engineering at system level, not just per call
• Evaluation methodology: designing test sets, measuring groundedness and accuracy, and building regression suites for non-deterministic output
• Guardrails and safety: prompt injection defence, output filtering, PII handling in prompts
• LangChain / LlamaIndex, with a clear view of when not to use a framework
• Tool and function calling: designing tool contracts, schemas, and error semantics
• Production experience with at least one agent framework: LangGraph, CrewAI, AutoGen, Semantic
• Orchestration and state management across long-running, multi-step workflows
• Agent memory: short-term context and persistent state
• Human-in-the-loop design: approval gates, escalation, and deciding where autonomy should stop
• Permissioning and sandboxing: scoping agent access to systems and data
• Failure recovery and loop control: preventing runaway execution and runaway cost
• Tracing and observability for agent runs (Langfuse, LangSmith, Arize Phoenix, or similar)
• MCP (Model Context Protocol)
• Multi-agent architecture, and the judgement to know when a single agent is the better answer
• Deep hands-on experience with at least one of:
• AWS: Bedrock, SageMaker, Lambda
• GCP: Vertex AI, Cloud Run
• Azure: Azure OpenAI, Azure AI Search, Azure ML
• Ability to become productive on the others, since client stacks vary
• IAM, secrets management, and cloud security fundamentals
• Docker, and enough CI/CD literacy to hand over cleanly
• Fine-tuning open-source models (LoRA, PEFT, QLoRA) and a view on when fine-tuning beats retrieval
• Lightweight AI/ML Ops exposure: enough monitoring to know when something has broken in production