We ship AI systems
that actually work.
The demo is the easy part. We build the document pipelines, internal search, and AI agents that survive production. In your cloud. On your data. Measured, not promised.
straight-through document extraction
to a prototype on your real data
revenue systems scaled at Zendesk
Senior engineers, not account managers · Your cloud, your code · NDA & DPA, signed first
- Document AI
- RAG & Internal Search
- AI Agents
- LLM Product Features
- Eval Harnesses
- Human-in-the-Loop
- AI MVPs
- Pilot Rescue
The problem
Every failed AI project started as a great demo.
Invoices retyped by hand. Claims packets read line by line. A chatbot pilot that wowed the boardroom in March and was quietly shelved by June. Most companies have both problems at once: manual work AI should be doing, and AI projects that never became systems.
The gap is almost never the model. It's the engineering around it: evals, validation, review queues, integrations, monitoring. That's the part we do.
cost to manually process one invoice. Automated: about $3.
of enterprise data is unstructured: documents, scans, PDFs.
of GenAI projects get abandoned after proof of concept, per Gartner.
What we build
AI systems with a measurable job.
One niche, done properly. Document AI is the flagship, and everything else ships with the same discipline. Evals first. Humans in the loop. Your cloud.
Flagship
Document AI
Pipelines that read your invoices, claims, and contracts: classify, extract, validate, route exceptions to human review, and land clean data in your ERP with an audit trail from every field back to the pixel. 95%+ straight-through.
02
RAG & internal search
Your wikis, tickets, and contracts answering questions with citations. Retrieval that's evaluated and tuned, not a vector-database demo.
03
AI agents & automation
Multi-step workflows with guardrails, checkpoints, and human approval where it matters. "Autonomous" shouldn't mean "unsupervised."
04
LLM features in your product
Structured outputs, eval suites, graceful fallbacks, and inference costs with a ceiling. Shipped inside your existing codebase.
05
AI MVPs
A production-quality first version of your AI product, built to be handed over, not rebuilt.
06
Pilot rescue
Your AI pilot stalled. We instrument it, measure what's actually broken, and finish the last 20% that keeps it out of production.
The flagship, dissected
A pipeline you can measure, not a magic box.
Document AI is a sequence of unglamorous, testable steps, and every one of them reports numbers: extraction accuracy, straight-through rate, exception rate, cost per document.
Ingest
Documents come in from wherever they already are: email inboxes, scanner folders, uploads, SFTP, your existing systems. PDFs, images, spreadsheets, even photos of paper.
Classify
Each document is identified and routed: invoice or credit note? ACORD form or adjuster's report? Multi-document packets get split into their parts before anything is extracted.
Extract
The right tool per document type: modern vision-language models, specialist OCR, or a cloud API when that's genuinely sufficient. Legacy OCR manages about 60% on handwriting. Not good enough. Every field comes out with a confidence score, not just a value.
Validate
Extracted data is checked against your rules and your records: does the PO exist, do the line items sum, is the IBAN valid? Anything that fails goes to a review queue where a person corrects only the flagged fields, and every correction becomes a test case that makes the pipeline better.
Integrate
Clean data lands where the work happens: your ERP, CRM, database, or API, with an audit trail from every field value back to the pixel on the page.
Use cases
Where this pays for itself first.
Finance
Invoices & AP
Header and line-item extraction, PO matching, GL coding suggestions. Fewer exceptions, faster close, early-payment discounts you actually catch.
Insurance
Claims packets
FNOL forms, ACORD forms, police reports, medical bills. Split, classified, and extracted so adjusters adjudicate instead of retype.
Logistics
Bills of lading & PODs
BOLs, proof-of-delivery, rate confirmations, customs docs. Shipment data in your TMS hours earlier, with fewer billing disputes.
Legal
Contracts
Parties, dates, renewal windows, obligations, non-standard clauses. Extracted into a register you can query instead of a folder you can't.
Lending
KYC & loan packets
Bank statements, pay stubs, tax forms, IDs. Faster decisions, consistent checks, an audit trail your regulator will like.
Different documents, same architecture. If it shows up as a PDF, it can probably be automated.
Receipts
Receipts, not promises.
Remote Node Labs is run by Alex Nikulin, the founder of CapyParse, a document processing platform for bookkeepers that turns financial and shipping documents into structured, confidence-scored data.
Before going independent, Alex spent four and a half years at Zendesk scaling the billing platform behind $1B+ in annual revenue for 200,000+ customers, and built search and segmentation systems at Punchh (acquired by PAR Technology).
shipping production software
annual revenue through billing systems scaled at Zendesk
verified records in a data product we built and sell
Ways to work together
Two ways to work with us.
We build it end-to-end
From sample documents to production: discovery, a labeled test set, the extraction pipeline, the review UI, integrations, monitoring, and documentation. You own all the code, and we hand it over properly.
Best when: the documents are piling up and you don't have an AI team to throw at them.
Book a callWe embed with your team
Your engineers get an AI specialist on the inside: architecture, model and vendor selection, eval harnesses, the hard 5%, code review. Meanwhile your team builds the skills to own it long-term.
Best when: you have strong engineers who just haven't shipped this particular thing before.
Book a callFAQ
Fair questions.
We handle sensitive documents. Where does the data go?
Wherever you need it to stay. Pipelines run in your cloud account (AWS, GCP, Azure) or on-prem; documents never transit our infrastructure. LLM calls go through deployments with zero data retention: your own Azure OpenAI, Bedrock, or Vertex endpoints, or self-hosted models when data can't leave at all. Nothing is used for model training. PII can be redacted before any third-party call, everything is logged for audit, and we'll sign your NDA and DPA before we see a single document.
How do you measure accuracy?
Against a labeled test set, not by eyeballing demos. Early on we label a few hundred of your real documents, the ugly ones included, and every change to the pipeline is scored against them: field-level precision and recall, straight-through rate, exception rate. You get numbers per document type and per field, so "95% accurate" means something specific: which fields, on which documents, at what confidence threshold. Fields below threshold don't get guessed; they get routed to a human.
Why not just use Azure Document Intelligence or Google Document AI?
Sometimes you should, and we'll tell you when a $0.001-per-page cloud API is enough. For clean, standard documents the prebuilt models are excellent, and a custom extractor would be a waste of your money. Where they fall short: unusual layouts, handwriting, multi-document packets, your validation rules, the human review workflow, and getting data into your ERP with an audit trail. The extraction API is maybe 20% of a working system. We build the other 80%, and often use those same cloud APIs as one component inside it.
What does "human-in-the-loop" actually look like?
A review queue, not a mailbox. Documents that fail validation or fall below the confidence threshold show up in a lightweight UI: the document on one side, extracted fields on the other, only the flagged fields highlighted. A reviewer corrects those, the record proceeds, and every correction is stored as a new test case, so the exception rate falls over time. Your people handle the 5%, not the 100%.
How long does it take?
Faster than you'd expect to something measurable, slower than a demo to something trustworthy. A prototype running on your real documents with honest accuracy numbers typically takes 2–4 weeks. Production usually lands in 6–12 weeks depending on document variety and your systems: integrations, review UI, monitoring, the long tail of edge cases. We'll give you a real estimate after seeing your documents, and not before.
How do we start?
Book a call and bring 10–20 sample documents (redacted is fine). In 30 minutes we can usually tell you what's automatable today, what accuracy is realistic, and roughly what it would take to build. If an off-the-shelf tool solves your problem, we'll say so and point you at it. The call costs you nothing either way.
Have an AI problem?
Show us.
Book a 30-minute call and bring 10 sample documents. You'll leave knowing what's automatable, what accuracy to expect, and what we'd build, even if our answer is "buy, don't build."