Thibaut Labarre replaced an expensive AWS Comprehend classifier with OpenAI prompts that could extract deeper information from news articles in a single day.
2
AngelList Relay parses investment documents and company updates, then organizes extracted terms and updates in an investor dashboard.
3
Thibaut Labarre favors off-the-shelf models for clear workflow improvements while reserving part of the team's time for custom models and new product capabilities.
Summary
Thibaut Labarre explains how AngelList moved from conventional machine learning to large language models for text-heavy investing workflows. His first system classified news articles so investors could follow portfolio companies, but AWS Comprehend imposed a fixed monthly cost and could not extract details such as acquirers and acquired companies. The team rewrote it with OpenAI prompts in a day. They then used the same approach to parse investment documents and company updates, which became AngelList Relay. Relay extracts terms, checks them against source documents, and gives investors a dashboard of their holdings and updates. Thibaut describes the engineering around prompts, including document formatting, chunking, human review, and regression testing. He also explains why AngelList routes requests through OpenAI and Azure OpenAI to increase capacity. His approach is deliberately focused: apply existing models to workflows that already require human labor, then explore custom models when the product case is clear.
AngelList started with a conventional classifier for company news
Thibaut Labarre joined AngelList after working at Amazon on natural language processing for customer feedback. One early AngelList project classified news articles and routed them to investor dashboards for the relevant portfolio companies. He built and deployed the classifier in about one or two months. AngelList had no data scientists or research scientists focused on machine learning, so Thibaut also developed the surrounding capability and later onboarded engineers to apply the same framework to other problems. The company gave employees substantial autonomy, which let him test the idea without a long approval process.
Off-the-shelf machine learning became too expensive and too shallow
The first system used AWS Comprehend and a custom text-classification model. It worked for a while, but its deployment model required at least one server at a fixed cost of about $1,000 per month. The classifier also could not answer more detailed questions about an article, such as who acquired a company or which company was acquired. Those limits made the system a poor fit for AngelList's needs. Thibaut says the team eventually replaced it with an OpenAI-based system in a single day, using prompts and a small number of examples rather than fine-tuning.
Prompts let the team expand the news workflow without rebuilding the system
The OpenAI system classified articles according to AngelList's categories and then extracted entities and events from the article text. New requirements could be added by changing the prompt. The team could distinguish company fundraising from fundraising by a venture fund, identify acquisition parties, and add other fields without rebuilding the model infrastructure. As GPT-3, GPT-3.5, and GPT-4 became available, the system improved without a corresponding model-training effort. The cost also shifted from keeping a server running to paying per request, which was cheaper for this workload.
AngelList uses source documents to check extracted information
For news and investment documents, the relevant text is supplied as context to the model. The system can check whether an extracted value actually appears in the source, which gives the team a direct way to catch unsupported output. Humans remain involved in some publishing and deduplication steps, especially when selecting the best article. Thibaut says the team gradually trusts the system with more complex tasks because the extraction results can be compared against source documents and structured records. This makes the risk manageable for workflows where the original legal or news document remains available.
Relay turns internal document parsing into an investor product
The team applied its document-extraction system to investment documents, extracting the company, investor, investment amount, valuation, cap, and other terms. It first helped AngelList operations review incoming documents and compare them with the information entered by investors. That internal workflow became AngelList Relay, a public product for investors, fund managers, and LPs. Users can forward investment documents and company updates to Relay. The product parses them, groups the information by company, and builds a dashboard that combines investment terms with later updates. Thibaut says the free tier supports five investment documents or company updates per month.
Prompt work includes document engineering and human review
Relay uses a sequence of prompts with LangChain. Incoming documents are classified first. Supported document types then receive a more advanced prompt that extracts terms and can inspect legal language for items that may need the investor's attention. Thibaut describes the process as simple from a system-design perspective, while saying that getting the prompts to produce the expected output took many iterations. The work also includes formatting documents, extracting text, handling long documents through chunking, and keeping a human in the loop. The team is still building a structured regression suite so prompt changes can be tested against known documents.
Domain experts should help write and improve prompts
Thibaut does not reduce the work to prompt writing. Engineers still need to prepare documents, run pipelines, deploy prompts, and operate the system at scale. At the same time, he wants lawyers and operations staff to tune prompts because they understand the business and legal domain better. Natural-language prompts lower the barrier for those employees to test changes and describe the output they need. Engineers can then focus on production concerns and scale. He expects this division of work to help AngelList move faster than requiring engineers to own every part of the prompt-development process.
AngelList prioritizes proven workflows before custom models
Thibaut sees a clear return from applying existing models to work that people already perform manually. AngelList's main effort is directed toward those workflows, with a smaller share of time reserved for new capabilities built from customer data. One possible future product would let an investor ask questions across their own investments, such as which seed investments later reached a Series B in a particular industry. Custom or open models could support more tailored experiences, but Thibaut says those efforts are still at the prototype and hackathon stage. The company does not want to add another model provider while OpenAI meets its current needs.
Capacity and vendor dependence are the main production constraints
The hardest operational problem has been getting enough access to GPT-4 and larger context versions, along with higher rate limits for document volume. AngelList now sends requests to both OpenAI and Azure OpenAI. Thibaut says Azure offers a more flexible cloud-style approach to capacity and has been more consistent for them, while using both services gives the system a fallback. The team is also considering models trained and deployed on its own hardware. That could reduce costs and give AngelList more control, but it would introduce its own operations work. For now, the team is prioritizing breadth of use cases over detailed cost optimization.
"We were able to kind of deprecate the whole thing and rewrite the whole system in a day and I'm not kidding, in one day, by leveraging large language models like OpenAI."Thibaut Labarre14:00
Who should watch
You are replacing manual document or text review and need a practical way to test whether an LLM can handle the workflow.
Your team is deciding whether to use a managed model API or build and operate its own models.
You need ideas for involving legal, operations, or other domain experts in prompt development without handing them production ownership.
