LandingAI

Extracting structured data from brokerage statements and tax forms. Fintech teams building portfolio tools spend most of their effort on ingestion. Holdings sit across several brokerages, and every statement arrives in its own PDF layout. Template-based OCR is the usual answer. It holds until a brokerage redesigns its statement, and then rules pile on rules. PortfoliMosaic drops the templates. The open-source assistant answers questions across brokerage documents in plain language. Extraction runs on Agentic Document Extraction (ADE) in two steps. ADE Parse returns clean markdown with tables intact. ADE Extract maps that markdown onto a fixed portfolio schema. Structured fields alone would not answer every question. Fee language lives in the document text. A holdings total is an aggregation. So a LangChain agent routes each question to semantic search, SQL, or a price API. The markdown then does double duty. That single parse pass feeds both the structured extraction and the retrieval index. The separation carries into any document pipeline. Worth reading with the author's own caveat. This is a reference implementation, with no authentication. Full write-up, source code, and video walkthrough in the comments.

Building an immigration financial audit pipeline on Amazon Web Services (AWS). Auditing 6 months of bank statements for one visa applicant takes 2 to 4 hours. Openomi, an audit platform for immigration consultants, does it in 60 seconds. It checks proof of funds against the rules of 10 distinct immigration programs. All of them run under Canada's IRCC. The hard part is program drift. Federal Skilled Worker requires CAD $15,263 across 6 consecutive months. Quebec Skilled Worker requires CAD $3,462 across 3. Canadian Experience Class requires no proof of funds at all. Mixing those rulesets produces false rejections. A Streamlit UI drops statements into S3. A Lambda orchestrator then calls two services in sequence. Agentic Document Extraction (ADE) runs first. ADE Parse converts each statement to markdown, holding multi-column layouts, handwritten notes and branch stamps. ADE Extract returns a typed schema: account holder, balances, currency, and every transaction. An AWS Bedrock Agent takes that JSON and runs the compliance and fraud checks. The core design decision was keeping program rules out of the system prompt. They live in a RAG knowledge base the agent queries before any document gets read. Full write-up, architecture, and source code in the comments.

Every document workflow comes down to two operations: parse and extract. Agentic Document Extraction (ADE) treats them as two distinct operations, and the split shapes how you build. Parse reads the whole document. You get Markdown plus a grounded document tree: every block carries a layout tag and line-level grounding, without any layout pre-training or fine-tuning on your side. Extract works on top of that parsed output. You define a schema naming the fields you want, and it returns each one as JSON, with a bounding box pointing back to where that value sits on the page. The sequence matters for cost. Parse once, keep the Markdown base, then run as many schemas against it as your field requirements evolve. Use Parse alone when downstream systems need visibility into the full document for search or retrieval. Bring in Extract when specific fields need to land in a database. Try ADE at ade.landing.ai

Your coding agent can now process documents without leaving the IDE. Add Agentic Document Extraction (ADE) skills to Claude Code, and the setup takes two commands. Connect to the plugin marketplace, install the ADE Document Processing skill, then reload your plugins. Once the skill is active, your agent parses documents into structured Markdown and pulls structured fields directly inside your workflow. No switching tools, no writing client boilerplate from scratch. Full setup steps live at https://lnkd.in/d9d_--7x.

The DeepLearning.AI Voice AI Hackathon wrapped on Saturday. Teams spent the day in Mountain View building voice agents on Sabre APIs and VocalBridge, and our team came to build alongside them! Ava Xia and Seshu Reddy built FinePrint: a voice interface for travel insurance policies. You upload a policy, ask it a question out loud, and the exact clause highlights on the page while the answer is read back to you in sync. The pipeline behind it is simple. ADE Parse turns the policy into page structure with bounding boxes, ADE Extract pulls typed benefit fields from that structure, VocalBridge carries the live voice, and Claude handles the clause-level reasoning. What makes it trustworthy is one strict rule: every spoken answer must resolve to a real line on the page. If the model invents a citation, it resolves to nothing and the answer never reaches your ears. The voice can only claim what the document shows. Want the code, docs, and a sample policy to try the full loop? Link in the comments

We're hiring a Customer-Facing Applied AI Engineer at LandingAI. The role sits at the intersection of engineering, product, customer success, and sales. You will design and deploy document AI pipelines for enterprise customers and shape the roadmap of Agentic Document Extraction. What you will do: → Own pre-sales technical engagements. → Design custom ADE demos and POCs. → Deploy pipelines on Snowflake, AWS, VPCs. → Lead workshops against hyperscaler tools. → Feed customer signal back into product. What we look for: → 5 plus years in AI or ML. → Track record in document AI or LLMs. → Two plus years on Snowflake or similar. → Strong Python to guide coding agents. → Ability to build with Claude CLI. → Enterprise sales fluency, MEDDIC. Application link in the comments.

Building an invoice-to-contract compliance pipeline. Enterprises pay overcharges that the contract was written to prevent. DocuFlow catches them in 45 seconds. DocuFlow is an invoice-to-contract compliance pipeline that flags pricing violations before payment goes out. A single invoice arrives in one of dozens of layouts, with merged cells, handwritten notes, and inconsistent labels. Its matching contract hides the pricing schedule on page 30, buried under nested clauses and cross-references. Reviewers cross-check line by line. DocuFlow uses Agentic Document Extraction (ADE) in two steps. ADE Parse maps the layout across tables, headers, and multi-column regions. ADE Extract returns the fields as structured JSON, from invoice line items to contract pricing caps and discount thresholds. Every clause lands in Postgres with its embedding and bounding box coordinates. When the engine flags Section 4.2.A on an invoice line, the dashboard links straight to that position in the source PDF. Measured on 500 invoice contract pairs: → 94% detection rate at 85% confidence → 4.2% false positive rate → Under 45 seconds per pair The core design decision was splitting extraction from comparison. ADE handled the messy to structured conversion. A deterministic rule engine handled the pricing check. Every violation stayed reproducible and audit-ready. Full write-up, demo video, and source code in the comments.

Every AI agent can act, but few can speak. The DeepLearning.AI Voice AI Hackathon on July 18 challenges builders to close that gap, specifically in travel, one of the most fragmented experiences out there. Flights, hotels, ground transport, dining, and experiences all live in different apps. None of them talk to each other. The challenge: pull all of it into a single conversation. One voice agent that books, manages, and adapts a full itinerary without switching apps or hunting through confirmation emails. Our CEO, Dan Maloney, joins the judging panel alongside Andrew Ng and others. Teams of up to 4. Application only, in person in Mountain View. $5,000 in prizes, product credits, and awards across categories. Submissions must use Sabre APIs and Vocal Bridge. 🗓️ July 18, 8 AM - 6:30 PM PT Apply: luma.com/vocalbridgeai

Type-safe document extraction with Zod and ADE. When you define an extraction schema as a plain JSON object, you get no TypeScript inference on what comes back. You end up casting the response or maintaining a separate interface alongside your schema. With ADE's TypeScript library, you can define your extraction schema directly with Zod. Convert it to JSON Schema with z.toJSONSchema(), pass it to the extract endpoint, and use z.infer to get the TypeScript type from the same schema definition. Here's an example using an invoice schema:

Most extraction stacks still run two engines to parse one document. One engine reads structured fields while another handles layout, a split that made sense when models struggled with complex tables and dense forms in the same pass. That split leaves teams maintaining two pipelines and reconciling their outputs before any downstream work begins. Agentic Document Extraction (ADE) reads the whole layout in a single call. The API returns every row, every subtotal, every adjustment code with its page and bounding box coordinates attached, so a reviewer can trace any value back to the pixel it came from. A multi-patient EOB puts this to work in one document, with claim blocks for six patients, five to seven billing rows each, subtotals per patient, and adjustment codes threaded through every row. ADE parses the entire page in one call and returns each field as a grounded chunk ready for the next stage of the pipeline. Try it in the Playground: ade.landing.ai