The Opportunity

This is a production AI engineering role building on a live platform. Solvace’s KAI copilot already runs multiple production agents (KPI analysis, record management, charts, FAQ, board analysis, meeting notes, greeting, record creator) serving manufacturing clients globally. The roadmap calls for a large library of specialised agents, multi-agent AI orchestration around operational workflows, voice interfaces, and MCP/A2A multi-agent communication.

This is not a research or maintenance role — it’s about building and shipping new AI agent capabilities on a platform that already has real users, real data, and real feedback loops. The broader engineering organisation is modernising the underlying application platform (migrating to .NET 8, containerising for Kubernetes, moving to PostgreSQL), and this role collaborates with that team where AI capabilities touch application data and infrastructure. But the primary mandate is forward-looking: expanding the agent library, building the agent builder platform, and scaling the AI copilot’s capabilities.

Core Technical Requirements

• Python — primary language; the agent engine is a FastAPI application. This is a Python-first team
• FastAPI — the production AI service runs on FastAPI; must be proficient in building, testing, and scaling FastAPI applications in production
• LangChain / LangGraph or equivalent — hands-on experience building agent workflows, tool-calling chains, and multi-step reasoning pipelines
• LLM APIs and model deployment — experience with Claude (Anthropic), OpenAI, or similar via direct API or abstraction layers. Must understand token budgets, context window management, and cost implications. Critically, must understand the trade-offs in managing context windows in high-throughput, large-scale LLM deployments — batching strategies, KV-cache management, and when to truncate vs. summarise context
• LLM fine-tuning and customisation — experience customising LLMs through fine-tuning (LoRA/QLoRA, instruction tuning) to create domain-specific models, and understanding quantization trade-offs (GPTQ, AWQ, GGUF) for balancing quality, latency, and cost in production deployments. Must be comfortable tuning model configuration parameters based on data and use case — temperature, top-p/top-k sampling, repetition penalty, frequency penalty, max tokens, stop sequences, and system prompt architecture — understanding how each affects output quality for different agent types (e.g., lower temperature for SQL generation agents vs. higher for creative meeting summarisation)
• LiteLLM or similar abstraction layers — the stack uses LiteLLM for provider-agnostic model access via AWS Bedrock
• RAG architecture — building retrieval-augmented generation pipelines: document chunking, embedding, retrieval, re-ranking, and response synthesis
• Vector databases — Weaviate (current stack) or equivalent (Pinecone, Qdrant, pgvector)
• SQL — agents query databases across multi-tenant schemas. The engineer must write and optimise SQL within agent tool definitions, with awareness of tenant isolation requirements
• Prompt engineering — systematic approach to prompt design, few-shot examples, system prompts, and output parsing
• Linux / CLI proficiency — must be comfortable working in Linux environments, containerised workflows, and command-line tooling

AI-Assisted Development Methodology

Solvace is transitioning towards AI-assisted development as a core engineering practice. This is especially important for the AI Engineer role:

• Hands-on experience with AI coding tools — Claude Code, OpenAI Codex, GitHub Copilot, Cursor, or similar. We expect this person to not only use these tools but to help define best practices for AI-assisted development across the team
• Spec-driven development — ability to write clear technical specifications that can be used to drive both human and AI-assisted implementation. Critical for agent development where prompt specifications, tool definitions, and expected behaviours need to be precisely documented
• Portfolio evidence — professional projects or side projects built with AI-assisted development tools are a must. Contributions to or experimentation with emerging projects like OpenClaw are a strong signal of someone who stays at the frontier
• Testing and evaluation rigour — experience building robust test suites and evaluation frameworks for AI systems, including non-deterministic outputs

Nice-to-Have

• Go or Rust — valued as evidence of strong backend engineering. Go is particularly relevant for building high-performance agent infrastructure
• AWS Bedrock — current LLM provider; experience with model invocation, inference profiles, and cost management
• SignalR / WebSockets — the copilot BFF layer uses SignalR for real-time meeting transcription streaming; understanding of real-time communication patterns
• C# / .NET — the BFF layer between the frontend and the Python agent engine is written in .NET 8. Ability to work across the full stack is valuable but not required
• Redis — used for session context and conversation memory
• Multi-tenant architecture — agents serve multiple clients with isolated data; experience with tenant-aware systems is a significant plus
• Cost optimisation for LLMs — per-token cost tracking is built in; experience optimising model selection, caching, and token usage
• Agent evaluation and testing — experience building test harnesses for non-deterministic AI systems
• Temporal or workflow orchestration — experience with workflow engines, multi-agent orchestration frameworks, or visual workflow builders is a strong plus

What You’ll Be Doing (First 6 Months)

1. Expand the agent library — build new specialised agents for roadmap capabilities including RCA virtual agent (leveraging rich historical root cause data with 5W2H analysis), expert suggestion agent, and virtual assessment/audit agent. Each new agent requires: tool definitions, prompt engineering, SQL query design, automated testing, and multi-tenant validation
2. Improve agent reliability and quality — implement error handling, fallback strategies, and response quality checks across existing agents. Build automated regression testing for agent outputs
3. Scale the meeting notes agent — the real-time transcription and summarisation agent needs hardening: handling longer meetings, multiple speakers, action item extraction accuracy, and integration with the action plan module for automatic follow-up creation
4. Build the agent orchestration foundation — prototype the workflow builder that allows non-technical users to compose multi-agent workflows around operational processes, set triggers, and schedule automated workflows. This is a greenfield product within a product
5. Optimise agent data access — work with the Data Engineer to migrate agent queries to optimised data layers (analytics platform, read replicas). Implement query caching, result caching, and intelligent fallback strategies
6. Collaborate with Research Engineer — implement prototypes (MCP server integration, advanced RAG patterns, voice interfaces) into production-grade, multi-tenant, cost-tracked agent code

Research vs. Applied

Strongly applied / production-focused. This person builds, ships, and operates production AI features. They should be able to read a research paper and understand it, but their primary skill is turning ideas into reliable, scalable, production-quality agent code. They’re measured on agents shipped and agent quality, not papers read.

Why Join?

• Major contribution to the copilot product — this is not a cog-in-the-machine role. They’ll join a team building an AI copilot platform that’s already in production with paying customers, with significant ownership over new agent capabilities and architectural decisions
• Production AI from day one — multiple agents already live, real users, real feedback. No 18-month R&D cycle before seeing impact
• Full-stack AI ownership — from prompt engineering to database queries to production deployment. No artificial boundaries between “ML” and “engineering”
• Agent orchestration platform — the multi-agent workflow builder is a greenfield product within a product; the right candidate gets to architect and build it from scratch
• Manufacturing domain — AI in manufacturing is early innings; the opportunity to define how AI copilots work on factory floors is genuinely novel
• Fast-moving team — the Innovation Hub operates at startup pace inside a funded company
• Career growth — as the AI team scales, early hires shape the team culture, architecture, and technical direction