Sr. GenAI Engineer

Steps AI · On-Site, Hyderabad, India · Full-Time

About Steps AI Steps AI is an agentic AI platform that powers customer-facing AI agents for businesses across e-commerce, Saa S, healthcare, real estate, Ed Tech, and financial services. Our agents go live in under five minutes from a single URL, learn the business by ingesting websites, documents, FAQs, and product catalogs, and deploy across web, messaging, and social channels. We are profitable, growing fast, and shipping production AI to customers every week. The Role We are hiring a Gen AI Engineer to own the agentic AI layer of our platform — the brains of every conversation our agents have. You will design and evolve our agent runtime, integrate new tools and channels, advance our retrieval and memory systems, and own end-to-end LLM observability. This role is AI-focused but not AI-only. You will also contribute to the surrounding backend services and data pipelines that power the AI layer. You will work alongside a peer Software Engineer focused on the platform side; together you will set the technical direction for major product areas.

Resumes will only be reviewed if the application form is completed.

Application form

What You Will Own

Agentic AI Systems (Primary)

• Agent runtime & control flow — design and evolve how our agents reason, route, recover, and hand off; build robust patterns for human escalation, partial-state resumption, and graceful tool failure
• Persona & behaviour engineering — shape how agents adapt across channels, business verticals, and brand voices; encode domain guardrails, retrieval mandates, and structured-output contracts
• Multi-provider LLM orchestration — abstract across leading model providers with per-persona routing, fallbacks, retries, and cost/latency tradeoffs
• Tool ecosystem — extend, debug, and ship integrations across CRMs, helpdesks, calendars, productivity suites, e-commerce backends, logistics carriers, messaging platforms, and internal services. A new provider should reach production in days, not weeks
• Retrieval & ranking — own the hybrid retrieval stack across vector and lexical signals, including reranking, recall/precision tuning, and ingestion-time quality controls
• Memory systems — long-term cross-session memory and short-term per-thread state, with clear contracts between them
• Streaming protocols — token-level streaming, intermediate event emission, interruption and resume semantics, and structured terminal events for client integration

Workflow Orchestration & Data Engineering

• Long-running workflows for ingestion, crawling, FAQ and Q&A synthesis, product discovery, sentiment work, outbound campaigns, and channel onboarding
• Queue & worker design — separation of heavy and light work, concurrency control, retry semantics, and end-to-end tracing
• Embedding & indexing pipelines — async batched writes, idempotent reindexing, provider rate-limit safety, and intermediate object-store staging
• Document understanding — multi-format parsing, OCR for scanned content, and structured extraction
• Web acquisition — adapter-based crawling with rate limiting and resilience; pluggable backends for varied site profiles
• Provider-agnostic embedding layer — designed so new embedding models drop in without touching downstream consumers

AI / Backend Integration You will not throw work over the wall. You will go into the backend stack when needed.

• The agent configuration pipeline that turns persona definitions into runnable behaviour
• New API endpoints powering AI-driven features — summarisation, FAQ synthesis, recommendation flows, and more
• Inbound webhook flows from commerce platforms and messaging providers that drive agent behaviour
• The plumbing that ties skill and tool configuration across our service boundaries

You Lead. You Don't Just Code.

• Own the AI roadmap for your area and propose what to build next
• Mentor and review the work of junior engineers as the team grows
• Run weekly architecture reviews with the founders
• Defend tradeoffs in writing through RFCs, design docs, and postmortems
• Be the on-call escalation point for AI / agent production issues
• Set the bar for code quality, testing rigour, and observability across the AI stack

Required Skills You should have shipped these to production and be able to demonstrate them live: Agent frameworks Production experience with Lang Graph, Lang Chain, or comparable agent runtimes — tool calling, state machines, checkpointing, streaming, human-in-the-loop LLM APIs Open AI, Azure Open AI, Anthropic, Google Gemini — direct integration and proxying patterns Vector databases At least one of Milvus, Pinecone, Weaviate, or Qdrant — schema design, dense + sparse retrieval, batched upsert, idempotency Embeddings & reranking Modern embedding models, dimensionality and quality tradeoffs, cross-encoder reranking Workflow orchestration Temporal, Airflow, Prefect, or equivalent at production scale Backend (Python) Fast API, async I/O, Pydantic, structured logging Backend (Type Script) Node or Nest JS — comfortable enough to ship full-stack features Observability LLM tracing (Langfuse or similar), Open Telemetry, error monitoring Cloud + Dev Ops Docker, AWS or GCP, CI/CD, production debugging

Strong Plus

• Multimodal model integration — vision and audio
• Speech: transcription and streaming synthesis
• Layout-aware document AI and advanced structured extraction
• Multi-agent patterns beyond single-loop architectures
• Advanced retrieval research — late-interaction models, learned sparse retrieval, and similar
• Open-source contributions to agent frameworks or vector databases
• Published papers, patents, or conference talks in the Gen AI / agentic space
• Domain expertise in e-commerce, govtech, healthtech, or financial services

Mindset We Hire For

• You think like a system architect, not a prompt engineer. Prompt-only "AI engineers" are not the right fit.
• You ship rapidly and iterate publicly. A new integration goes from idea to production in days.
• You read source code instead of asking Stack Overflow. When a framework behaves unexpectedly, you read its internals.
• You measure what you ship. Token usage, latency, retrieval quality, regression rates — all instrumented from day one.
• You take ownership of failures. When a customer reports a bad outcome, you trace it through prompt, retrieval, and tool output, then fix the root cause.