Overview

We are building AI systems that read, understand and act on real business documents, bank statements, financial reports, policy documents and forms and putting them into production where accuracy and cost both matters.

Responsibilities

• Select, deploy and benchmark open-source LLMs and vision-language models for specific, narrow use cases not general chat.
• Build evaluation sets from real documents and define what "good" means numerically (field-level accuracy, extraction recall, hallucination rate) before shipping.
• Run structured comparisons between models and approaches, and write up the trade-offs so the team can make a decision.
• Apply quantization, batching and other optimizations to fit models into a sensible GPU budget.
• Design multi-step pipelines that combine deterministic code, ML models and LLM calls and know when not to use an LLM.
• Optimize for latency, cost and reliability: caching, batching, request routing, fallback tiers, retries and graceful degradation.
• Instrument pipelines so failures are visible and traceable rather than silent.
• Package models and services with Docker, expose them behind clean APIs, and deploy them to our GPU and CPU infrastructure.
• Handle the unglamorous production concerns: cold starts, timeouts, concurrency limits, versioning, rollback and monitoring.
• Own on-call-style responsibility for the AI features you build, including cost tracking.

Qualifications

• Strong Python: type hints, async/await, dataclasses/Pydantic, clean module design, testing.
• REST API development with FastAPI (or Flask/Django with a willingness to move to FastAPI).
• Git, code review discipline, and the ability to write code someone else can maintain.
• Comfortable in Linux and on the command line.
• Working knowledge of PyTorch and the Hugging Face ecosystem (transformers, tokenizers, accelerate).
• Understanding of inference-time concepts: tokenization, context windows, batching, precision (FP16/BF16/INT8), memory footprint.
• Ability to read a model card and a paper well enough to judge whether a model fits a use case.
• Hands-on experience with at least two of: pypdfium2, PyMuPDF, pdfplumber, pdfminer.six, Docling, Unstructured, Surya, DocTR, LayoutLM family.
• Practical OCR experience (Tesseract, PaddleOCR, or a cloud OCR) and an understanding of when OCR is the wrong tool.
• Experience extracting tables from PDFs and dealing with merged cells, multi-line rows, and inconsistent column layouts.
• vLLM, TGI, Ollama, llama.cpp, or Triton Inference Server.
• Quantization formats and tooling: GGUF, AWQ, GPTQ, bitsandbytes, ONNX Runtime, INT8 export.
• Serverless GPU platforms: Modal, RunPod, Replicate, Baseten including cold-start and container-lifecycle management.
• LoRA / QLoRA fine-tuning with PEFT for narrow, task-specific improvements.
• Practical use of open VLMs: Qwen2.5-VL, InternVL, Granite Vision, Molmo, Phi-Vision, or similar.
• Awareness of where VLMs hallucinate especially on numeric and financial content and patterns for constraining them (using the model for layout only, sourcing values from the text layer, constrained decoding).
• Workflow orchestration: Dagster, Airflow, Prefect, or Temporal.
• Async job patterns: Celery, RQ, or platform-native spawn/poll patterns.
• LLM orchestration frameworks (LangGraph, LlamaIndex, Haystack) with the judgement to know when plain Python is a better answer.
• Structured output enforcement: Instructor, Outlines, XGrammar, JSON schema / tool-use modes.
• Building golden datasets and regression suites for extraction tasks.
• Eval tooling: promptfoo, DeepEval, Ragas, or in-house harnesses.
• LLM tracing and monitoring: Langfuse, Arize Phoenix, LangSmith, OpenTelemetry.
• Rule engines and policy evaluation (Open Policy Agent / Rego, Drools, rule-engine).
• Experience in fintech, lending, insurance or accounting documents.
• Handling of PII and data-security practices in document pipelines.
• Contributions to open-source ML or document-processing projects.
• Real production ownership from month one your work goes to actual users, not a demo.
• Genuinely hard technical problems in document AI, not wrappers over an API.
• Small team, short decision cycles, direct access to leadership.