Staff Engineer (GenAI)

Location: Bengaluru, India

Department: Engineering

Experience: 8+

About Zenskar

Funding

The Problem We're Solving

What Customers Say

"We're saving 200+ hours/quarter on invoicing and receivables by completely automating our recurring billing."

— Noy Kalansky, Finance Controller, Pontera

"Zenskar automates revenue recognition accurately for our value-based billing: agents reducing manual hours by 70%."

— Matt Barnard, VP Finance, Vertice

"Zenskar’s agents automated 90% of our billing, integrated with our CRM, and accelerated revenue collection by a month."

— Ming Lui, VP Finance, Yembo

"Sardine had spent 4 years running billing in-house for high-volume, usage-based pricing. Zenskar took care of it all."

— Sardine team

"We launched our product 4 months faster instead of building an in-house system for our usage-based pricing."

— Kshitij Gupta, CEO, 100ms

About the role

What you'll do

• Turn ambiguous product asks into working LLM-based features: prompt and context design, tool calling, orchestration across model calls, structured output handling
• Design for the failure modes non-deterministic systems actually have: hallucination, drift, silent wrongness. Build guardrails, and know the difference between demo-grade and production-grade because you've shipped the gap between them
• Make sure an LLM is never the unverified source of truth for a financial decision. Design the verification or human-in-loop layer wherever money is actually at stake
• Pick the right model or vendor for the job (cost, quality, latency, data policy) and avoid lock-in. Once chosen, run it efficiently: token economics, caching, batching, per-request cost tracking
• Set the technical standard other teams build their Gen AI features against, and get multiple teams to actually adopt it without having direct authority over them
• Build and own the eval harness that catches quality regressions before they ship, not ad hoc spot-checks
• Know when not to use an LLM at all, where a deterministic system would be cheaper, more reliable, or simply correct
• Name and quantify technical debt in Gen AI systems (prompt sprawl, eval gaps, untracked cost, brittle integrations) and negotiate time to address it
• Shape who joins the team by holding a technical bar in interview loops, without owning headcount decisions

What drives you

• You've seen AI systems fail in production and got obsessed with understanding why
• You think about reliability, evaluation, and observability before you think about model selection
• You believe probabilistic systems deserve the same engineering rigor as deterministic ones
• You care about systems users can trust, not systems that look impressive in a demo
• Gen AI's baseline moves monthly for you, not yearly. You actively track it: specific people, sources, releases, and can point to a real decision that changed because of a recent shift

What you'll need

• 8+ years of software engineering experience, including meaningful experience shipping AI-powered products to production
• A real, shipped-at-scale system in your track record, doesn't have to be Gen AI. If it's ML work, it shipped to real users, not fine-tuning or research that stayed in a notebook
• Deep understanding of LLM application architecture: tool use, structured outputs, retrieval, orchestration, and where these actually break in production
• Experience building agentic systems that do multi-step reasoning and interact with tools or business systems
• A track record of treating prompts as versioned, testable, deployable artifacts, not strings scattered through the codebase
• Strong RAG fundamentals: retrieval quality, chunking, embeddings, evaluation, knowledge system design
• Experience designing evaluation frameworks and regression pipelines for AI systems
• Strong grasp of AI observability: tracing, monitoring, feedback loops, production debugging
• A track record of getting more than one team to adopt a pattern or standard you set, without owning those teams
• Judgment for when not to use an LLM at all
• Strong backend engineering skills, and enough frontend ability to own an AI experience end to end
• Can clearly walk through an AI system you've built two levels below the pitch: what failed, what you learned, how you improved reliability over time

Good to have

• Memory architectures: episodic memory, procedural memory, retrieval systems, knowledge stores
• Agent orchestration frameworks (LangGraph, Pydantic AI, OpenAI Agents SDK, CrewAI, or a custom runtime)
• Long-running autonomous workflows and event-driven agent systems
• Voice, multimodal, or real-time AI systems
• Fine-tuning experience and understanding of model internals beyond API consumption
• Open-source model deployment and inference infrastructure
• Experience with financial systems, billing platforms, revenue operations, accounting, or fintech
• Familiarity with MCP, tool ecosystems, and AI platform architecture
• Startup experience and comfort operating with high ownership and no formal authority

Location

• Hybrid - 2 days per week
• Office Location: Indiranagar, Bengaluru.
• Address: 3rd Floor, A wing No 1, Carlton Towers, HAL Old Airport Rd, HAL 2nd Stage, Indiranagar, Bengaluru, Karnataka 560008.

Interview process

• R0, Recruiter screen (30 min): Fit, motivation, and a quick check that you have real, specific experience behind your background, not generic answers.
• R1, Technical: critique a broken Gen AI integration (60 min): You'll be handed a small billing-shaped Gen AI integration with seeded defects and asked to find them, including the one that would go unnoticed until it was already wrong on a customer's invoice.
• R2, Technical: design a new integration and its eval framework (60 min): An intentionally underspecified product ask. You design the integration, decide whether an LLM belongs in it at all, choose a model with real reasoning, and design how you'd know it's working before a customer complains.
• R3, Project audit (60 min): A numbers-anchored walkthrough of one real system you've shipped at real scale, architecture, usage, cost, and the debt it accumulated.
• R4, Architect and influence (60 min): Whether you've actually set a technical pattern that more than one team adopted, and how you've held a hiring bar without owning headcount.
• R5, Bar raiser (60 min): Values, judgment, and a deeper probe on whatever came through thinnest earlier in the process.
• Reference checks: Two former colleagues, run by two different people on our side, including a deep dive on a Gen AI system you shipped and what happened after it hit production.

How to apply