AI CFO agent stack
Proof-of-concept agents across order-to-cash, payables, close, consolidation, treasury and B2B payments. Rules make every decision; Claude only writes the memo, email or status note.
AI & Data Science
Agents that do real finance work (collections, close, treasury, cash forecasting) with rules that make the decisions, models that write the explanations, and a person who signs off. Code, experiments and notes from building them.
build log
Proof-of-concept agents across order-to-cash, payables, close, consolidation, treasury and B2B payments. Rules make every decision; Claude only writes the memo, email or status note.
An agent for large code migrations. Claude rewrites one function at a time; static checks, property-based differential tests and SMT equivalence proofs decide what a human needs to review.
Small runnable demos of the core idea in Chain-of-Thought, ReAct, Toolformer, Reflexion, Tree of Thoughts, Voyager, AutoGen and three more. Runs offline, or live against Claude.
A reference implementation of a Netflix-style foundation model: one autoregressive transformer over full interaction histories, with sparse attention and metadata embeddings.
Reads open invoices and customer history, picks the next best action per account, does low-stakes ones itself and sends the rest to an approval queue, with a full audit trail.
Month-end Excel workbooks in, a guided evidence-first review out: the agent walks the CFO through the numbers and ends with a management-ready deck narrative.
Categorising 200 bank statement lines two ways, and what it says about which treasury labels you can post without a human checking them.
Agent at work
Cash dips to $232k, $18k under the $250k buffer. Recommend moving two vendor payments ($40k) from week 7 to week 9.
writing
There are many parts of the CFO vertical can be enabled by AI to increase the productivity and at the same time improve the per...
An agent I watched last quarter did everything right and still got the answer wrong. It read the request, produced a clean five...
How modern AI applications coordinate models, tools, memory, and workflows to build production-grade intelligent systems
Nothing matches that search.
One model. Every member’s full history. Many recommenders.
An agent I watched last quarter did everything right and still got the answer wrong. It read the request, produced a clean five...
Why understanding how large language models think may become one of the most important breakthroughs in artificial intelligence.
There are many parts of the CFO vertical can be enabled by AI to increase the productivity and at the same time improve the per...
Anthropic’s latest essay suggests we may have crossed an important threshold: AI is no longer just helping humans work faster —...
Why the future of data governance is not just about helping humans understand data, but helping AI understand what your busines...
How modern AI applications coordinate models, tools, memory, and workflows to build production-grade intelligent systems
The lending industry is standing at the intersection of two powerful forces: rapid AI-driven innovation and increasing regulato...
Notes on Booking.com's framework for choosing ML best practices by linking each one to measurable quality attributes, so teams ...
Nearly 90% of models never reach production. The common reasons, from shifting requirements during development to deployment an...
How DeepSeek, a Chinese AI lab, shipped frontier-level models at a fraction of competitors' cost, and why it shifts the AI deba...
Notes on "MemGPT: Towards LLMs as Operating Systems": giving LLMs tiered memory, paged like an OS, so they can handle long conv...
How multimodal LLMs take in images as well as text, the main architectures for combining them, and a summary of recent research...
A quick tour of LangChain: what it does, how it connects LLMs to external tools, context and memory, and the features that make...