About the Role
As a Senior AI Engineer, you will design, train, and ship the models that turn messy, semi-structured business documents into reliable structured data — for example extracting and decomposing the goods, parties, and terms from trade-finance documents and Letters of Credit. This is a hands-on, data-centric role: you own problems end to end, from pulling and cleaning production data, to training custom on-prem models and standing up LLM-based extractors, to proving they work with honest evaluation and deploying them into a live microservice platform.
You will set the technical approach for extraction workstreams, be relentless about data quality and evaluation rigor, and act as a trusted technical partner to the product and domain teams who depend on the results.
Key Responsibilities
Build production extraction models for complex documents: token classification / sequence labeling, span segmentation and decomposition, and structured attribute extraction from semi-structured text (SWIFT/LC fields, invoices, packing lists, certificates).
Fine-tune and train small, on-prem transformer models (encoder token-classifiers, small language models) on GPU, and stand up LLM-based extraction (Azure OpenAI / GPT) as baselines and PoCs — with anti-hallucination grounding and structured output.
Own the data: pull from production stores (MongoDB), establish label provenance, deduplicate, and build leakage-free splits (e.g., by transaction/tenant) so metrics mean something.
Build and maintain evaluation harnesses: report honest precision / recall / F1, make results reproducible, and actively hunt for contamination, overfitting, and mislabeled data before quoting a number.
Make and defend trade-offs across accuracy, precision-vs-recall, cost, and latency, matched to how the model is actually used in production.
Package models as containerized GPU/CPU microservices and ship them through GitOps (ArgoCD / Helm / Kapitan) across dev, QA, and production, with the right config and secrets handling.
Define standards for evaluation, monitoring, and model lifecycle (MLOps / LLMOps), and build review tooling that lets domain experts validate and correct model output.
Partner with product and domain experts to translate business rules and human-labeled data into robust, precision-first extraction; mentor engineers and set best practices.
Required Experience & Skills
5+ years of professional AI engineering, including production ML/NLP systems.
Strong Python and applied deep learning with PyTorch and Hugging Face Transformers — real experience with token classification, sequence labeling, and fine-tuning encoder models.
Hands-on model training on GPU (CUDA, mixed precision, long-context), and comfort operating remote training boxes (SSH / Teleport, Azure or GCP).
Practical experience with LLM extraction pipelines (Azure OpenAI / OpenAI): structured output, grounding to source text, and prompt design used as a baseline, not the finish line.
Data-centric ML: dataset construction from raw stores (MongoDB), deduplication, leakage-aware train/test splitting, and label-quality analysis.
Evaluation rigor: metric design, honest reporting, and an instinct for when a number is too good to be true (leakage, memorization, sparse or synthetic labels).
Deployment fundamentals: Docker, Kubernetes, and GitOps (ArgoCD / Helm / Kapitan a strong plus).
Clear communication of results, trade-offs, and uncertainty to both engineers and non-technical stakeholders.
Nice to Have
Domain exposure to trade finance, banking documents, or Letter-of-Credit processing.
Document AI / OCR experience (hOCR, Azure Document Intelligence), table extraction, layout analysis, and document classification/routing.
Serving/orchestration of ML services (durable queues, model facades) and cost/latency optimization for GPU inference.
RAG / vector search where it genuinely fits — treated as one tool among many, not the default answer.
