AI Engineer in India, भारत

Certa (certa. ai ai) is an AI-native third-party risk management platform that unifies the full third-party lifecycle, from onboarding through continuous monitoring and offboarding. Enterprises from the Fortune 500 to fast-growing startups use our no-code platform to manage vendors, suppliers, and partners across risk, compliance, and ESG. Trusted by Honeywell, Uber, MARS, and Boeing, we’ve processed over 10 million entities for 100,000+ users across 120+ countries. At Certa, AI agents do real work for enterprise risk and procurement teams: Studio designs workflows, Data Agent answers questions over customer data, and Risk Agent assesses third parties. The model is the same one everyone can buy - the harness is what makes ours do the job reliably, safely, and cheaply, and you’ll own that harness. This isn’t a “build AI features” role: you own an agent end to end and you’re measured on whether it does the task - success rate, latency, cost per task, and what customers stop doing by hand. It’s a small team with high ownership, where you’ll work directly with PMs, design, and customers, and be trusted to make the call. WHAT YOU’LL DO - Own the agent loop: Step control, planning, retries, error feedback, subagents, and parallel tool calls for long, multi-step tasks. - Design tools & actions: Build the tools, MCP servers, and skills our agents use - narrow, well-described, permissioned, and safe to call against customer data. - Engineer context: Decide what the model sees on every step - compaction, memory, retrieval-as-a-tool, and how agents read large enterprise records without drowning.

- Build evals that gate every change: Golden sets, regression suites, and trace-based scoring wired into CI, so no prompt, model, or harness change ships blind. - Own model choices & the runtime: Route across models on capability, latency, and cost, running new model launches through the eval suite and re-tuning the harness when the frontier moves, with durable sessions, sandboxed execution, nested-trace observability, token budgets, and graceful fallbacks for multi-tenant enterprise customers. - Own the backend systems: Design services end to end (data model, APIs, storage choice, failure modes), build async pipelines (queues, workers, retries, idempotency, backpressure), and make them fast and cheap at scale (caching, batching, indexing, p95/p99 latency budgets) on multi-tenant AWS with isolation, least-privilege access, PII handling, and audit trails. - Own it in production: Write the design doc, weigh the trade-offs, defend them in review, and own the service in production. WHAT SUCCESS LOOKS LIKE - Day 30 - Backend: Own one Studio backend service end to end and ship a production change (a new API, a pipeline fix, or a performance improvement) with a measurable result, such as a lower p95 latency or error rate, and join its on-call rotation. - Day 60 - Fix: Find the top failure mode of an existing agent from its traces and ship a fix that lifts its task success rate by a measurable amount on our eval set (target: +10 points). - Day 90 - Build: Ship one new agent capability or backend service you designed from scratch, live with at least one customer, with a written design doc and a target metric you set and hit. WHAT YOU’LL NEED - Backend & system design: 3+ years building production backend systems with strong system design - you’ve designed and scaled services under real load, picked the data stores, modeled the data, and handled async processing, queues, caching, idempotency, and failure in production. Python preferred. - Shipped an agent: You’ve shipped at least one LLM agent to real users and can explain its loop, its tools, its failure modes, and the numbers that showed it worked. - Evals that change decisions: You’ve built evals that changed a decision - you killed, shipped, or reworked something because the eval said so.