AI Engineer

LITSLINK is looking for an experienced LLM Architect to work on a complex production-grade document analysis and structured data extraction system.

We are looking for more than an LLM Engineer who implements predefined technical tasks. We need a hands-on architect who can design an ML/LLM system end-to-end, justify architectural decisions, evaluate technical trade-offs, and ensure the system is accurate, reliable, efficient, and scalable in production.

We value deep knowledge not only of LLMs and RAG, but also of NLP, Information Retrieval, data modeling, evaluation, and production ML systems — including an understanding of underlying approaches and algorithms, their trade-offs, and when to use them.

About the project

We are building a complex LLM-powered agent that analyzes documents and extracts predefined structured data from them.

Three properties are critical:

• Accuracy — extracted data must be faithful to the source documents, including exact references and quotes.
• Validity — generated structures must conform to predefined schemas, data types, required fields, and business rules.
• No duplicates — the system must not recreate structures that have already been accepted or rejected.

Your main responsibility will be to design the architecture of the extraction pipeline and contribute hands-on to the implementation of its key components.

What you’ll do

• Design the end-to-end architecture of an LLM-powered document extraction pipeline.
• Determine where LLMs should be used and where deterministic algorithms and executable validation are more appropriate.
• Design multi-pass pipelines such as extract → validate → deduplicate/link → verify → iterate.
• Build measurable quality frameworks using golden datasets, Precision/Recall/F1, regression testing, and automated quality gates in CI.
• Design hybrid exact + semantic retrieval, indexing, and reranking strategies.
• Design identity-based deduplication, record linkage, and entity resolution approaches.
• Design incremental re-indexing, versioning, and supersession logic.
• Ensure pipelines are idempotent and resumable, with appropriate caching and token/cost optimization.
• Define executable acceptance criteria for architectural decisions.
• Contribute hands-on to the implementation and validation of critical technical components.

Must-have

• Experience shipping 3+ LLM-powered systems to production, each with a measurable quality/evaluation framework you personally built.
• Hands-on experience with golden sets, Precision/Recall or similar quality metrics, regression testing, and CI quality gates.
• Strong practical knowledge of LLM systems, NLP, and Information Retrieval.
• Experience with structured document/information extraction.
• Proven ability to separate model-driven extraction from deterministic validation in executable code.
• Hands-on experience with RAG, embeddings, semantic search, hybrid retrieval, reranking, and vector databases.
• Strong Python and software engineering background.
• Experience with data modeling, identity-based deduplication, record linkage, and/or entity resolution.
• Experience designing complex multi-stage / agentic pipelines.
• Strong understanding of production concerns such as reliability, performance, observability, scalability, and cost optimization.

What is especially important to us

We need someone who can do more than recommend a technology — you should be able to explain and defend the architectural decision behind it.

For example, saying “we should use Chroma for vector search” is not enough. We expect you to explain why Chroma is appropriate for the specific use case, how it compares with pgvector or other alternatives, what trade-offs each option introduces, and under what conditions you would choose a different solution.

We expect the same depth of reasoning around embeddings, retrieval strategies, chunking, reranking, validation, deduplication, evaluation, and agent architecture.

We are looking for both an architect and a practitioner — someone who has made these decisions in real production systems and can also validate and implement the proposed architecture hands-on.

Nice to have

• Experience working with insurance, finance, legal, or other regulated-domain documents.
• Experience with event-sourced or lakehouse-style record stores: append-only, versioned, time-travel.
• Knowledge graph construction.
• Advanced record linkage / entity resolution experience.
• Experience with adversarial-verification agent patterns.
• Experience building custom LLM evaluation suites/tooling when off-the-shelf frameworks were not sufficient.

Engagement

• Remote
• Full-time for the first 2+ months
• Transition to part-time afterward
• Direct communication with the technical team and client
• Long-term cooperation

This is not a role focused primarily on standard chatbot/RAG integrations or prompt engineering. We are looking for someone with deep production experience who can design a robust ML/LLM system, justify key architectural decisions, and define measurable ways to prove that the system works correctly.

Відгукнутись на вакансію