We’re seeking a Senior AI Engineer to design and ship production-grade agentic AI systems that automate complex workflows end-to-end. This is a hands-on, senior role with significant technical ownership. You’ll work closely with the Chief Architect, product, engineering, and domain experts to translate ambiguous, high-impact problems into reliable AI-driven user experiences.

What Success Looks Like

Ship AI capabilities that measurably improve user outcomes (quality, time saved, throughput)

Build systems that are reliable by design: evals, observability, safety, and cost/latency controls from day one

Iterate quickly using a tight loop of instrument → evaluate → improve → deploy

What You’ll Do

Agentic AI Feature & Workflow Development

Build and integrate AI-driven features using LLM APIs (OpenAI / Azure OpenAI, Anthropic, Gemini on Vertex AI)
Design and implement tool-using agents (structured function calling, schema validation, retries, fallbacks)
Build multi-agent workflows when appropriate (e.g., planner/worker, reviewer/critic, specialist routing)



and know when a simpler architecture is better
Create agentic workflows such as document understanding, extraction, reasoning over evidence, task automation, and multi-step decision support
Own context engineering end-to-end:
agile context assembly (retrieval + state + tool outputs)
context budgeting and compression/summarization
grounding strategies to reduce hallucinations and improve consistency
Implement retrieval-augmented generation (RAG) and search workflows using off-the-shelf vector stores and embedding services
Evaluation, Quality & Iteration (Core)

Establish evaluation frameworks for accuracy, reliability, and output quality
Build task-specific eval suites: golden datasets, adversarial cases, regression tests, and rubric-based scoring
Set up automated evaluation pipelines and release gates (CI/CD-friendly) tied to prompt/model/version changes
Define and monitor online metrics (e.g., task success rate

📌 Senior AI Engineer (Pune)
🏢 Onit India
📍 Pune