Hassan Alam, Founder and CEO of AI Asset Management, joins AI InterConnect to talk about a part of AI that gets far less attention than chatbots: documents. AI Asset Management turns complex documents such as PDFs into structured data that modern AI systems can use. Hassan's argument is simple: the quality of the context you give a language model decides how well it works.
Two startups in document AI
Hassan has worked in data and document AI for a long time. His previous two startups, both exited successfully, extracted content from documents using rule-based natural language processing.
The lesson that led him here: the better the quality of the data going in, the better any downstream AI method works. Today the downstream models are language models and agents, but the principle is the same, whether a model is small or large.
The spreadsheet test
A simple experiment showed him the size of the problem. He gave an image of a spreadsheet table to both ChatGPT and Claude. One missed two of 22 columns; the other missed five of 50 rows. The models have improved since, but extraction quality still decides the outcome.
He also points to context rot: as you add more context, even to a model with a large context window, performance starts to degrade. Load ten large documents into a chatbot, he says, and you will see it get confused. Two contracts with different vendors in the same format are enough for a model to mix up the details.
Structure is information
Documents encode meaning in their structure: who the author is, what is a table or a caption, and how information flows. Using that structure lets a system read only what matters, such as an abstract, the results or the sources.
AI Asset Management reconstructs each document as a knowledge graph, at the level of paragraphs and document objects rather than individual named entities, so each piece keeps its context. When a query arrives, the system finds the right nodes, expands the graph and passes clean context to the language model of the customer's choice.
Domain models and research rigour
General-purpose retrieval systems treat every document the same, but financial, legal and research documents look very different. Hassan builds a general model and tunes it for each domain, which he says is faster and more accurate.
His work has been supported by research grants, including from DARPA, and he collaborates with researchers at MIT and Cambridge, which brings rigour to turning fundamental research into production systems. He also sees automation as the key shift, compared with expensive manual labelling services such as Scale AI.
Secure, efficient and small
Many organizations have private documents they do not want shared, but cannot afford to retrain a language model; Hassan notes that even Bloomberg concluded it could not. AI Asset Management acts as a preprocessing step that keeps documents secure and lets a company run inference on a few GPUs in its own data center.
With clean, well-structured data, even small language models perform much better, which opens the door to edge devices. His goal is for someone to scan a legal document on a phone and get a reliable answer on the spot.
Advice for builders
Hassan's advice is to find the 80/20 case, where most of the value lies, and make the system robust there. AI is probabilistic and will never get everything right, so keep humans in the loop and use AI to raise their productivity.
He relies on a diverse team and on measuring twice before implementing once. And for any founder, he says, be prepared for a long journey with many ups and downs, and cherish good team members; many of his have followed him to this third company.
Watch the full conversation
Watch Hassan Alam explain how structured documents and knowledge graphs make language models more accurate.
AI InterConnect hosts conversations with founders, investors and operators building with AI. Explore more episodes on the AI InterConnect podcast page.