About This Role

8 to 18 years of experience

Skills: GenAI, LLM Application Development, RAG, Multi-Model Experience

Key skills –

- GenAI / LLM Application Development
- Prompt Engineering
- Multi-Agent Architecture
- Multi-Model Experience (GPT, Claude, Gemini, etc.)
- AI Solution Architecture
- RAG (Retrieval-Augmented Generation)
- AI Guardrails & Governance
- Function Calling / Structured Outputs
- AI Evaluation & Testing

What you'll be doing

- Design the multi-agent architecture— how an agent (system prompt, allowed model(s), validation rules, output shape) is defined, versioned, and selected at runtime, so adding the next agent to Client AI Studio doesn't mean re-architecting the platform.
- Integrate multiple models behind a common interface, and make deliberate model-selection calls per task rather than defaulting to the biggest model available — e.g. Claude Sonnet for content drafting that needs judgement, a smaller/faster model for mechanical, high-volume, well-specified transforms . You'll need to be able to justify these choices, not just make them.
- Own prompt engineering across every agent— the Jira Creation Agent's story/AC/DoD generation, the KCI Notification Agent's content drafting per Medium (Email/SMS/Letter), and the "regenerate this one field" and "tell the AI what's wrong" refinement patterns used consistently across agents. Iterate based on real output quality, not first-draft-and-ship.
- Define the AI service layer's contracts— the APIs the front-end and back-end developers integrate against ("generate," "regenerate field," "validate," "check duplicate," etc.)



— working with the back-end developer, who owns the broader application backend, auth, and persistence.
- Build guardrails, not just prompts: required-field validation, duplicate detection, and — critically — knowing where a deterministic system should do the work instead of an LLM (e.g. final HTML assembly for compliance-critical output, where consistency matters more than generative flexibility).
- Establish how we evaluate agent quality— test prompts, example sets, and a way to tell whether a prompt change made things better or worse before it ships, rather than relying on spot-checking.
- Work with the testerto define AI-specific test cases that sit outside normal functional QA — prompt injection resistance, degenerate/ambiguous inputs, hallucination checks — since this isn't covered by testing a deterministic feature.
- Work with content and compliance stakeholdersto encode what the AI is and isn't allowed to touch (e.g. legal footer text that must never be silently edited), so this is enforced by design rather than caught in review.
- Document each agent— system prompt structure, model choice and why, known limitations — so agents stay maintainable by the team,



not a black box only you understand.

What we're looking for

Essential

- Production experience building LLM-powered features — not prototyping in a chat window, but designing prompts, structured inputs/outputs, and failure handling for something other people depend on.
- Experience working across multiple models/providers and making real cost/latency/quality tradeoffs — able to explain why a given task uses one model over another.
- Comfortable owning an API contract and working as the AI specialist inside a team where front-end, back-end, and testing are separately owned — this is not a full-stack role.
- A clear sense of where an LLM is the wrong tool — recognising when deterministic logic or templating is more reliable than generation, and designing accordingly.
- Able to work directly with non-technical stakeholders (content writers, compliance, product owners) and turn "this doesn't feel right" into concrete prompt or architecture decisions.

Desirable

- Experience building an agent/skills framework (tool use, function calling, structured system-prompt libraries) rather than a single-purpose chatbot.
- Experience in a regulated or compliance-sensitive setting, where "the AI drafted something slightly wrong" has real consequences.
- Basic retrieval/RAG experience, for agents that need to reference existing content (e.g. an existing KCI to clone from, Jira project configuration) — doesn't need to be a full vector-search implementation, but the pattern should be familiar.

📌 AI Engineer (Bengaluru)
🏢 Bounteous
📍 Bengaluru