Applied AI Engineer - Document Intelligence

Fortegra is building AI-enabled systems for commercial insurance. Our document-intelligence software processes emails, scanned PDFs, policy packets, loss runs, and spreadsheets with inconsistent layouts. It identifies what was received, extracts and reconciles relevant facts, preserves links to the source, and presents the results for review and action. Current work includes underwriting submission triage and audits of issued policy packets.

Reliability is central to this work. A field can look plausible and still be wrong. A parser can misread a page without raising an error, or a workflow can omit a document entirely. We need to find those failures, understand their impact, and keep the system dependable as document formats, models, prompts, and parsing tools change.

We're hiring an engineer to build and improve these production systems. You'll work closely with the technical lead for applied AI on architecture, data representation, and model choices. You'll take bounded features from a business question or production failure through planning, implementation, evaluation, and production readiness. You'll surface decisions that need broader input, and take responsibility for larger features and workflows as you learn the system and domain.

WHAT YOU'LL WORK ON

- Build applied-AI product features. Work across document parsing and OCR, classification and routing, structured extraction, normalization and reconciliation, source grounding, retrieval and summarization, and the deterministic logic around them. Most candidates will bring depth in some of these areas and an interest in working across their boundaries.

- Drive the evaluation loop. Inspect real documents, traces, and model behavior; maintain representative evaluation sets; classify failures; choose the next change from the evidence; and requalify the result. Outputs must be complete, support material claims with source evidence, and reconcile related values.

- Shape features with domain experts. Work with underwriters, auditors, claims professionals, and other specialists to clarify definitions, exceptions, source authority, and acceptance criteria. Turn those decisions into schemas, examples, validation rules, and product behavior.

- Compare and choose technical approaches. Test models, OCR and parsing tools, managed services, and internal implementations against the same representative documents. Recommend an approach based on quality, cost, latency, reproducibility, failure modes, and operational overhead. Build for the current problem, then reuse patterns that prove durable.

- The day-to-day work is production engineering: building features, debugging strange behavior, reviewing code, improving tests and observability, and operating services in a Python and Azure Functions codebase.

EARLY PRIORITIES

In your first months, you'll improve one existing capability through evaluation and requalification; ship a bounded document-intelligence feature; and complete a model, parser, or service comparison with the tests, artifacts, and handoff needed to continue it. Your scope will grow as you learn the system and domain and demonstrate sound judgment.

HOW WE WORK

- Choose the tool based on the problem. Use deterministic code for stable rules and LLMs for variable interpretation. Tests, evaluations, and production evidence support release decisions.

- Work effectively with coding agents. Claude Code, OpenAI Codex, Cursor, and similar tools are part of daily development. Give agents useful context and verifiers, choose how much autonomy to grant, and remain responsible for architecture, review, testing, and what gets merged.

- Follow failures through the system. A defect may involve documents, prompts, model behavior, application logic, and product expectations. We investigate across those boundaries and preserve concise plans, findings, and handoffs.

WHAT WE'RE LOOKING FOR

- Production software engineering. You've built and shipped Python services or product features and can reason about testing, observability, reliability, security, cost, and latency.

- Applied-AI judgment. You understand common LLM failure modes and how context, data quality, evaluation design, and model or API choices affect results. You've worked hands-on with structured outputs, tool calling, context design, or retrieval.

- Evidence-based debugging. You inspect source artifacts, traces, ground truth, evaluation sets, structured error analysis, tests, and production evidence to determine what happened and whether a change helped. Come ready to walk us through one example.

- Ability to shape incomplete work with domain experts. You can turn an underspecified problem into a plan, learn what its fields and rules mean, make progress, and identify decisions that need broader input.

- Practical technical judgment. You can choose among deterministic code, schemas, rules, search, product changes, and LLMs based on the problem.

- Responsible use of coding agents. You can delegate implementation, provide context and verification, choose an appropriate level of autonomy, and review, test, and correct the result. Experience with our exact tools is optional.

- Nice to have: document processing, OCR, information extraction, retrieval, or search experience; insurance or another regulated, document-intensive domain; Azure, serverless, or distributed processing; evidence, provenance, citation, or human-review systems.

- This production-software role focuses on dependable outputs and real failures. Model-training research and notebook-based analytics are outside its day-to-day scope.