Bonus add-ons and learning companions

1Why "bonus"

The business case in Part B is built on the core alone: ingestion, retrieval, the chatbot and the API. The add-ons below are not counted in the ROI. Each runs on the same governed knowledge, so switching one on does not mean another copy of our documents, another vendor or another permission model. If any of them replaces a tool we would otherwise buy or build, that is extra value on top of B6.

2Content Studio: documentation, manufactured

A document-authoring agent. Given a short brief, it runs an Author → Revise → Format pipeline grounded on our own knowledge base and source files, and produces a finished, published document with an evidence trail.

• Grounded, not invented: every document is built from our corpus, so it says what our sources say.
• Proven at volume: Trilivy's own library of 111 technical documents (~318,000 words) was produced with it. The ecosystem review grades it A-, 8.7.
• One configuration file controls layout and generation, so every document follows our house style.
• Persona Splitter: reframes the same document for executive, technical and adopter readers.

Use it for: policy summaries, onboarding guides, product FAQs, RFP answer libraries, internal knowledge articles. When good documentation becomes cheap, the question changes from "what can we afford to document?" to "what do we choose to leave undocumented?"

3Course Maker and the Workshop LMS: training, manufactured

Course Maker turns a course title and a short description into a complete course:

1. We enter a title and description, and optionally source notes.
2. The AI proposes a full syllabus: modules, lessons, topics, quizzes, assessments and hands-on exercises.
3. Our course owner reviews and approves it.
4. The AI writes the content into the LMS, where it can be edited and published.

The Workshop LMS around it handles the catalogue, enrollment, progress, hands-on submissions and certificates. It already ships with 84 lessons in five role tracks (C-level, product owner, technical team, adopter, strategist) that train our own people on the platform. Finished courseware is exported back into the knowledge base as Q&A, so what we teach also becomes something the chatbot can answer.

4Mobile apps: a jumpstart for iPhone and Android

Two fully working cross-platform apps, built in Flutter, for iPhone and Android:

Both come with source code. They are templates we can rebrand, customize and connect to our own data and workflows, instead of starting a mobile project from an empty folder. For comparison, Part B3 estimates $60k–$150k to build a hybrid app and $250k–$400k for a native one.

5Talk to a Person: human handoff

AI should not be the end of the road. When a user needs a human, the conversation goes to our staff with its full context:

• The user asks for a person, with a button in the chat or by saying so in their own words.
• A ticket opens in the Handoff Inbox, carrying the conversation so far, so nobody has to ask the user to repeat themselves.
• Our operators work the queue: take a ticket, see what is theirs and what is unassigned, and reply.
• Email notifications go to the instance's operators and to a configurable address, and staff replies reach the user by email.
• Configured per assistant in its "Talk to a Person" settings, so an internal IT bot and a customer-facing bot can route to different teams.

This matters most for customer-facing use (application 6 in Part A) and for regulated topics where an AI answer must be confirmed by a person.

6Learning companions

Two standalone teaching tools, not in the production pipeline and not graded. Nothing depends on them. They are included because the biggest predictor of AI program success is whether the people running it understand what they run.

The open-source and self-hosting case

This companion answers the question every board eventually asks: "What if the AI provider's price triples, the terms change, or our data cannot leave the building?"

• A self-hosted path already runs alongside the commercial APIs. Open-source models such as Llama run through Ollama on dedicated servers, behind the same provider boundary as every other model. Switching is a configuration change.
• Economics: self-hosting can reach 60–80% cost savings at scale for high-volume workloads, and keeps data on hardware we govern.
• Honest trade-off: open models in Ollama typically have a 6,000–8,000-token context for the whole prompt. Large 128K contexts need a heavier serving stack such as vLLM.

The Micro-LLM

A complete language model small enough to understand: a 2-layer transformer trained on ~50 financial question-and-answer pairs that fits on a laptop. Every stage (tokenization, training, generation) is documented and hands-on, so our engineers and decision-makers see how a model predicts text, why context length matters and where hallucination comes from.

7What our team should probe