AI Legal Assessment Workflow

On Jan 30, 2026, at 09:34, Dariusz Jan Nowak <dariusz.pgry@gmail.com> wrote: Thank you for your valuable feedback. The platform will operate on a subscription basis, so it will be available to users (an office writing for others and the users themselves—companies and individuals). I'm currently building the backend without a database connection. I'm testing the logic based on source documents. Each application is assessed using formal and substantive criteria. The AI evaluates the application and then wants to build the assessment made by human experts and teach the AI this assessment as case studies. There are many dependencies in the case of eligibility. Currently, I have a case where the AI assessed the application in the same way as the human, but a provision in the law (referring to a government act) confirms the incorrect assessment. I will protest the assessment. Will the AI try to challenge me? I'll add web scraping to it, and if it finds a provision referencing regulations (laws, regulations, EU directives, etc.), I'll provide feedback. Here, I'll be building other functionalities, such as content anchors (def get_citation_info) and an optimizer: improving metrics, risk assessment, cost rationalization, etc. I'm not thinking about the Holy Graal yet :), which is built-in AI in Excel (I can have extensive business plans for funding applications). I'll build it into an Excel file based on the AI data, fill it in, create a JSON, and load it into the RAG engine – to put it simply. We can also implement Dynamic Persona MoE RAG architecture. I'll also take a look at openNotebook, but for now, this is all I can show you: <image.png> See also what this guy from Germany did: https://www.chateic.com/ Regards Dariusz pt., 30 sty 2026 o 14:00 Daniel Kliewer <danielkliewer@gmail.com> napisał(a): Hi Dariusz Based on what’s publicly available, the NotebookLM MCP does give you a solid, well integrated core for document grounded reasoning, but it intentionally stops short of deep verification, optimization, or custom evaluation workflows. If your platform must strictly adhere to grant documentation and justify outputs, I would treat NotebookLM MCP as an augmentation layer rather than a foundation: host and structure your documents locally, add your own verification and evaluation pipeline, for example via a lightweight knowledge graph or rule based checks, and use the MCP call as a reasoning engine rather than a source of truth. That approach trades some efficiency for control and auditability, which is usually the correct trade off for grant writing systems. Rebuilding everything from scratch is possible, but in my experience it’s slower and harder to match NotebookLM’s baseline quality, wrapping it with your own verification logic is the more pragmatic path if correctness matters more than raw speed. The Dynamic Persona MoE RAG architecture extends standard retrieval augmented generation by routing both retrieval and reasoning through a mixture of explicitly defined personas rather than a single monolithic context. Each persona encodes a constrained perspective, skill set, and evaluation bias, allowing the system to decompose complex tasks into parallel, interpretable reasoning paths. Retrieved evidence is selectively weighted and cross checked across personas, with disagreements surfaced and reconciled through an evaluation layer. This makes reasoning more robust, reduces hallucination by design, and allows verification effort to scale dynamically with task complexity rather than being fixed at generation time. In an ideal world you could just build your own equivalent or use an open source version like OpenNotebook and build around that. Maybe you could try that. Build a wrapper around OpenNotebook that adds adding your own knowledge graph and evaluations likevero-evaluations are what I have used before and had good results. Actually this would be a great way for me to verify my new Dynamic Persona MoE RAG architecture. I could either build around OpenNotebook and try to maintain data sovereignty altogether, which honestly is what you should do if you want to offer your services to people who require more security, but it would allow me a way to test what I have been working on. Either way, I would explore OpenNotebook and Vero-Evaluations and see if adding your own local wrapper might solve the problems you are having. Daniel On Jan 30, 2026, at 05:32, Dariusz Jan Nowak <dariusz.pgry@gmail.com> wrote: Hello Daniel My name is Dariusz, from Poland. I saw your post on reddit about: "How I Built a Fully Uncensored, Persona-Driven AI Chatbot Using MCP and NotebookLM" I'm currently building a new web platform for writing grant applications; I'm using Google's AIStudio for this. It's crucial to read the competition documentation carefully and base all platform functionality on it. I've been using NotebookLM so far, but it lacks verification, optimization, and other tools. I'm wondering whether to continue building or use the NotebookLM MCP? What do you think? Best regards, Dariusz

Here is a concise one paragraph reply you can send. It avoids dashes and stays factual and grounded. :::writing{id="83421" variant="email" subject="Re: Platform architecture and verification workflow"} Hi Dariusz, what you describe is exactly the class of problem where a layered architecture performs better than a single reasoning engine. An AI should not challenge you autonomously but rather surface conflicts between human assessment, statutory provisions, and precedent as explicit, inspectable artifacts. Scraping and anchoring primary legal sources, attaching structured citations, and feeding corrected outcomes back as case studies is the right direction, especially if you want auditability and protest readiness. I would treat NotebookLM or similar systems strictly as document grounded reasoning components, not arbiters of correctness, and keep eligibility logic, legal precedence resolution, and optimization metrics under your control. Your Excel to JSON to RAG loop is pragmatic and aligns well with enterprise workflows, and a Dynamic Persona MoE RAG layer could add real value by separating legal compliance, risk assessment, and optimization into distinct evaluators whose disagreements are visible rather than averaged away. Overall, you are building toward a defensible system rather than a persuasive one, which is exactly what grant and regulatory contexts require. Best regards, Daniel :::