Senior Applied AI Engineer
AI Summary
Novara provides safety and operational risk management software that empowers organizations to identify and resolve issues before they become incidents. Through the Flex and Risk Management Center platforms, Novara helps organizations address operational risk proactively by unifying data, increasing workforce engagement, and proactively managing risk.
About this role
Position Description:
Knowledge, Experience, Requirements:
• SaaS Product Experience: 5+ years of software development experience, with at least 2 years spent
building and scaling production-grade AI features in a cloud-native SaaS environment.
• Educational Background: Strong academic background in Computer Science, Data Science, Software
Engineering, or a highly quantitative field (e.g., Mathematics, Physics, Statistics). Bachelor's degree in
Computer Science, Engineering, or a related technical discipline preferred.
• Technical Stack: You must have hands-on, production experience with the following technologies:
- Languages: Python (Expert/Senior level), TypeScript/JavaScript (Strongly Preferred).
- AI Frameworks: LangGraph, LangChain, Vercel AI SDK or equivalent.
- AWS Infrastructure: Amazon Bedrock, ECS Fargate, S3, SQS, EventBridge, KMS, AWS Lambda, Amazon
Comprehend, IAM.
- Databases & Search: PostgreSQL / pgVector, Amazon OpenSearch Serverless, SQLAlchemy.
- Data Processing: Pandas, NumPy, PyPDF, Layout-OCR engines.
- API & Protocols: REST, Server-Sent Events (SSE), Webhooks, and Model Context Protocol (MCP).
• Hands-on AWS Background: Strong experience designing secure AWS architectures using Least Privilege IAM execution roles, SigV4 API signing, and KMS envelope encryption.
• RAG at Scale: Experience indexing and searching datasets scaling into millions of document chunks, with a proven understanding of Direct Bulk Indexing APIs.
• System and Security Architecture: Solid understanding of authentication patterns (OAuth 2.0, JWT passthrough) and how to isolate data logically in multi-tenant shared databases.
• Clean Code Advocate: Demonstrated ability to write clean, unit-tested, and well-documented Python
code, utilizing self-correction loops and graceful degradation patterns to handle model latency and API
rate-limiting limits.
• Collaboration & Agile: Strong communication and collaboration skills, thriving in an agile, team-based environment.
Nice-to-Haves: The following experience will be highly valued:
• Machine Learning & Predictive Modeling: Practical experience training and serving classical ML models (e.g., Isolation Forest, One-Class SVM, or unsupervised clustering) for behavioral baselining, anomaly detection, or predictive risk scoring.
• Experience developing React-based micro-frontends or canvas-style Generative UI layouts.
• Contributions to the open-source Model Context Protocol (MCP) ecosystem.
• Background in EHS (Environmental Health & Safety) or ESG (Environmental, Social, and Governance)
software systems.
• AWS Certified Machine Learning – Specialty or AWS Certified Solutions Architect – Professional.
Success Criteria:
• Core AI & Orchestration - Key expectations for AI platform engineering:
• Agentic State Machines: Design and implement complex, multi-agent state machines and stateful
graphs using LangGraph and LangChain to support autonomous decision-making and self-correcting
loops.
• Dynamic Agent & Workflow Registries: Architect database-driven registries (using PostgreSQL) to
dynamically discover, load, and configure agent definitions, system prompts, and task workflows at
runtime without redeploying code.
• Optimized LLM Routing: Build intent-based routing engines that evaluate user queries and direct them to either deterministic execution layers (e.g., Python code interpreters running over in-memory DataFrames) or semantic retrieval layers (RAG).
• Observability & Cost Tracking: Configure centralized telemetry pipelines and AI Gateways for token tracking, caching, rate limiting, and real-time streaming of internal graph execution traces (via ServerSent Events).
• Advanced RAG & Data Engineering - Key expectations for data pipelines and search systems:
- Production-Grade RAG on AWS: Build and maintain a dual-engine vector search architecture: Amazon OpenSearch Serverless for unstructured policy, regulation, and SOP document retrieval, and PostgreSQL + pgvector for structured transactional logs, incident histories, and audit records.
- Serverless Ingestion Pipelines: Build scalable, event-driven ingestion pipelines using AWS S3, SQS, EventBridge, and AWS Fargate to parse raw documents (PDF, Word, CSV) into Markdown.
- Context Preservation & Visual RAG: Implement advanced chunking strategies, including slidingwindow paragraph overlaps, header breadcrumb perpetuation, and Vision Transformer (ViT) visualenrichment models to summarize embedded charts, diagrams, and stamps.
- Automated PII Redaction: Integrate Amazon Comprehend or custom LLM classifiers inside the Fargate worker container to scrub names, emails, and SSNs before data is indexed.
• EHS Integration & MCP - Key expectations for security, integration, and guardrails:
- Model Context Protocol (MCP) Servers: Build standardized, decoupled MCP servers that wrap legacy REST APIs (Java/C# backends), exposing databases, schemas, and actions as dynamically discoverable tools for the AI agents.
- Prompt-Independent Security (RBAC): Implement user-delegated token pass-through (JWT forwarding) so that data-access permissions are enforced mechanically by the legacy API. Design hard metadata filters (where tenant_id = jwt.tenant_id) in OpenSearch and pgvector to ensure multi-tenant isolation.
- HITL Write Guardrails: Configure Human-in-the-Loop (HITL) state breakpoints in LangGraph to halt write-mutations, broadcasting the pending action to administrators for UI-based approval.
• Generative UI Layouts - Key expectations for UI integration:
- Dynamic Component Rendering: Design structured JSON widget schemas (representing tables, Recharts graphs, checklists, and forms) generated dynamically by the backend agents to enable zerostate rendering of layouts in the Next.js UI.