AI Engineer Agentic Software Development (Client Site)

About the Role

Software development itself is being rebuilt around AI. You'll build the agents that do it - embedded inside a financial institution's own engineering organisation.

As a forward-deployed AI Engineer you sit with the client's tech team and build agents that their engineers use every day to run their software development lifecycle: requirement review architecture planning coding review deployment debugging. Your users are professional engineers working on real systems which sets a high bar - an agent that produces plausible-but-wrong output gets abandoned in a week.

This demands strong software engineering fundamentals. You can't build an agent that reviews architecture or writes production code unless you deeply understand how architecture and production code actually work. Equally you need a real point of view on where AI genuinely helps in the SDLC - and where it doesn't.
You'll report into Vecton's central engineering team with dotted-line accountability to the client's technical leadership.

What You'll Build

Agents across the client's SDLC for example:

Requirement agents - review BRDs/PRDs for gaps ambiguity and completeness of the proposed solution
Architecture agents - high-level solution architecture security architecture application-level LLD
Planning agents - decompose a project into precise executable development tasks
Coding agents - pick up those tasks and write test and iterate on the code
Review agents - code and PR review against the client's standards patterns and security expectations

Deployment agents - pipeline execution release checks environment handling
Diagnosis agents - pull logs traces and metrics via the right tooling to isolate and debug failures
And the layer that makes them work:
Context orchestration - designing how each agent gets the right context at the right time: codebase insight (repo indexing observability/Grafana data) third-party integration docs the client's internal standards prior architectural decisions
Harness decisions - evaluating build-your-own vs. established agent SDKs (Claude Code Agent SDKs and similar) and owning that call with clear reasoning
Skills tools and references - building the tool definitions skills and reference material that move agent output from "plausible" to "actually correct" for this client's stack
Evaluation - defining what "good output" means for each agent and measuring it
Beyond agentic SDLC

You'll also build AI-centered applications for the client's business use cases - LLM-backed workflows RAG systems and multi-agent solutions running inside their infrastructure and compliance boundaries.

What Being Forward-Deployed Actually Means

Learn the client's stack repos CI/CD ticketing coding standards and release process well enough to encode them into agents
Work directly with their engineers and tech leadership to find where agents will actually save time - and where they won't
Build inside their constraints: on-prem or VPC-hosted models data residency code that can't leave their network audit requirements
Drive adoption - demo onboard gather feedback iterate; usage is the success metric not delivery
Contribute to solution proposals effort estimates and PoC demos alongside Vecton's central team
Feed patterns components and learnings back into Vecton's knowledge base
What We're Looking For
3–6 years of strong software development experience - you've designed shipped and maintained production systems not just prototypes
Solid fundamentals in system design API design testing version control workflows CI/CD and debugging production issues

Python proficiency for AI/ML workloads plus working knowledge of at least one full stack (MERN/MEAN or Java/Spring Boot)
Hands-on with LLMs prompt engineering RAG and agentic frameworks (LangChain LangGraph CrewAI AutoGen or equivalent)
Practical experience with vector databases and embedding-based retrieval (Pinecone Weaviate pgvector)
A working thesis on how AI is changing software development - where it genuinely accelerates work where it fails and why
Daily user of AI-assisted development tools (Claude Code Cursor or equivalent) - you should be able to describe what they do well and where they break down
Ability to hold your own with senior client engineers and architects and comfort with ambiguity - much of this doesn't have an established playbook yet