> Markdown version of [/jobs/ext/4213634-ai-engineer](https://www.wearedevelopers.com/jobs/ext/4213634-ai-engineer). Every page supports `.md` or `Accept: text/markdown`. Links point to the HTML versions so they work for humans too. Agent guide: [/agents.md](https://www.wearedevelopers.com/agents.md). --- # AI Engineer - **Company:** Tessera Labs - **Location:** San Jose, CA, United States - **Experience:** Experienced - **Contract:** Permanent contract - **Skills:** Application Programming Interfaces (APIs), Artificial Intelligence, Static Program Analysis, Distributed Systems, Graph Database, Python (Programming Language), Oracle (Applications), Role-Based Access Control, Salesforce.Com, SAP (Applications), Single Sign-On, TypeScript, Supervised Learning, Data Processing, Enterprise Software Applications, Feature Engineering, Retrieval-Augmented Generation, Large Language Models, Snowflake, Database Optimization, Prompt Engineering, Model Validation, Juniper, Workday, Mulesoft, Servicenow - **Published:** October 31, 2026 - **Apply:** https://www.careerbuilder.com/job-details/ai-engineer-san-jose-ca--d8ae4ec1-884a-4514-bc3c-725894f405e5 ## About the Role Artificial Intelligence (AI), Enterprise Applications, Artificial Intelligence (AI), Data Processing, SAP, Salesforce.com, Oracle, Sales, Juniper Networks M-Series, Venture Capital, * Have 3+ years building and operating production software, with meaningful recent time on LLM-powered systems. * Have shipped agentic systems that real users depend on, and have owned the incident when they broke. * Are fluent in the current agent toolkit - tool calling, orchestration, context engineering, RAG, evals, tracing - and can say where each one stops working. * Have built retrieval systems against messy real-world corpora, and have measured them rather than assumed them. * Have real traditional ML in your background - supervised learning, feature engineering, model evaluation - and reach for it when it beats a prompt. * Treat evaluation as engineering rather than a report you generate at the end. * Think in systems and customer outcomes, not model metrics. * Are comfortable with non-determinism and have opinions about building reliably on top of it. * Write strong Python and are comfortable in TypeScript. * Ship fast, and verify before you claim it works. * Find large, ugly, undocumented enterprise systems interesting rather than beneath you. That instinct is most of the job, and it's rarer than it should be. Strong candidates may also have * Hands-on experience with enterprise platforms, their APIs, and their extension models - SAP, Salesforce, Workday, Oracle, Snowflake, MuleSoft, ServiceNow. * Experience with knowledge graphs, ontologies, or semantic models over messy real-world systems. * Experience with code analysis, program transformation, or automated refactoring at scale. * Experience with MCP, sub-agents, or agent skill/plugin architectures. * A background in distributed systems or workflow engines, where partial failure is the default case. * Experience at an early-stage startup or as a founder, where scope was whatever needed doing that week. * Familiarity with enterprise security and compliance realities: SSO, RBAC, segregation of duties, PII handling, SOC 2, data residency., We're looking for someone who enjoys building reliable, scalable systems that help teams move faster. If you take pride in cutting-edge AI, value clear ownership, and want to have a meaningful impact in a fast-moving environment, you'll fit right in. ## Description * Design and ship the production agents that do transformation work: understanding a landscape, planning a change, executing it across process, data, and code, and proving it was correct. * Build the tool layer - typed, permissioned, well-documented interfaces that let agents read and act across enterprise systems without ever exceeding what a human approver authorized. * Improve agent performance through prompting, context construction, tool-use strategy, and decision logic - and know which of those a given failure calls for. * Build the retrieval layer over enterprise artifacts and metadata: chunking and indexing strategies for content that doesn't resemble prose, hybrid search, reranking, grounding, and the evaluation that tells you whether any of it helped. * Manage context deliberately. Enterprise artifacts are enormous and a forty-step run accumulates history fast; deciding what the model sees, what gets compressed, and what gets dropped is a first-class engineering problem here, not a prompt detail. * Use classical ML where it's the right tool. Plenty of the work inside an agent pipeline - routing, ranking, classification, anomaly detection, confidence estimation - is better served by a small supervised model than by another LLM call, and knowing the difference is part of the job. * Build the eval and monitoring layer: task sets drawn from real customer landscapes, regression coverage on every deploy, and alerting that catches a changed schema, a revoked authorization, or a new model version before the customer does. * Instrument every run so each model call, tool invocation, decision, and human approval can be reconstructed afterward. This is what "passes audit" means in practice, and it's a product requirement rather than a nice-to-have. * Diagnose production failures from execution traces down to root cause, then close the gap in the system rather than in a single prompt. * Design the guardrails, approval gates, and rollback paths that make an agent safe to point at a live enterprise landscape. * Build capability that generalizes. Anything that only works for one customer is a bug in the product, not a feature of the engagement. Representative projects * Shipping the agent that lets a customer retire a third of their custom estate in a quarter - and the review surface that makes a human comfortable approving each call in minutes rather than days. * Designing the approval and write semantics for agents operating in a live landscape, so nothing reaches production without a traceable human decision and no partial failure leaves systems inconsistent. * Building the reconciliation agent that makes two merged companies' data agree, plus the eval set that tells us when it's confident and when it's guessing. * Standing up the replay-evaluation pipeline that scores a candidate agent version against a month of recorded runs - real system responses, mocked writes - so a regression is caught before anything is touched in a customer environment. * Taking a pattern that worked at one customer and turning it into a platform primitive that works at the next four without a human rewriting it.