ML / AI Engineer

About the role

Compliance teams spend most of their week shoving PDFs into folders. Tracegence flips that: a document lands, our pipeline OCRs it, classifies it, extracts the structured fields, validates them against rules + history, and surfaces what's about to expire, what doesn't match, and what's worth a human's attention. Three AI tiers do that work today: Document AI Form Parser at the OCR boundary, Vertex AI (Gemini 2.5 Flash) for classify + extract + chat, and a predictive layer that scores renewal risk and recall exposure. There's a mock path at every boundary so we can develop and CI without burning Vertex quota. You'd own how this layer evolves: prompt design and eval, when to fine-tune vs prompt, switching the fraud detector from mock to a real SageMaker-deployed model, and bringing real RAG validation online (Vertex embeddings + retrieval). The interesting work is in the eval - "is this CoA's water-activity field consistent with the supplier's last 12 months?" is harder than it looks.

What you'll do

• Own the AI pipeline end-to-end: prompts, model choice, mock paths, eval suites
• Design and ship the next iteration of classification, extraction, and validation accuracy
• Stand up RAG validation (Vertex embeddings + retrieval) properly - move it off RAG_VALIDATION=off
• Replace the SageMaker fraud-detection mock with a real deployed model
• Build offline eval datasets from real production traffic so model swaps don't regress silently
• Work with backend on guardrails: retries, timeouts, fallback paths when Vertex degrades

What we look for

• 3+ years shipping ML systems in production - not just notebooks
• Comfortable with prompt engineering AND classical ML; can tell when each is the right tool
• Hands-on with at least one of: Vertex AI, AWS SageMaker, OpenAI/Anthropic APIs in a real product
• Can read a transformer paper and decide if it changes your roadmap
• Comfortable writing Python services that other engineers will read and modify