01 How it works
Follow one piece of information from the source to my screen.
The walkthrough plays LADON, then HERETIC, then the two working together, and stops. Select any tab or step to jump to it.
What Happens
Why It Works This Way
This is an illustration of the design. The agents are still being built, so there are no alerts, notes, or performance figures to show yet.
02 Why it is built this way
Four decisions that shape everything else.
Match the Tool to the Type of Data
This is the most important rule in the design. Sending everything through a language model is slow, costly, and unreliable with numbers. Each kind of data gets the method that handles it best.
| Type of data | Examples | How it is handled |
|---|---|---|
| Structured | Prices, ratios, financial statement figures | Code and tools that calculate exactly. No model is asked to do arithmetic. |
| Semi-structured | SEC filings with known sections | Find the right section first, then read only that part closely. |
| Unstructured | News, commentary, management language | A language model, which is what it is good at: reading and summarizing text. |
Two Agents with Different Jobs
Watching and discovering pull in opposite directions. Watching needs to be fast, frequent, and quiet unless something matters. Discovering needs to be slow and thorough. One agent asked to do both does neither well, so LADON is built for vigilance and HERETIC for depth. When HERETIC finds a name worth following, it hands it to LADON.
Local and Private
Both agents run on local models on my own computer, inside an open-source, self-hosted AI workspace. My holdings and research never leave the machine, and there is no usage bill, which matters on a student budget.
The Agents Research. I Decide.
Neither agent places trades or gives a final verdict. LADON tells me what changed and shows the source. HERETIC lays out the case for and the case against. The judgment stays with a person, which is how I think this technology should be used in wealth management.
03 Where it stands
Working foundations, agents still in progress.
- Done: a custom tool server with 14 finance tools, connected to the workspace and confirmed working.
- Done: a local model running on the laptop's graphics card, with a one-click script that starts the whole system.
- Learned the hard way: the tools were silently not reaching the model at first. The workspace's tool-selection step depended on a database that was not running, so it fell back to a default set that left mine out. Finding that took longer than building the tools.
- Also learned: model choice matters more than expected. Several models could not call tools reliably, and a free hosted option ran into output limits, which is what pushed the project to local models.
- Next: LADON's scheduled checks and alert format, then HERETIC's filing-reading workflow.
What I am taking from it. Most of the work in an AI system is not the AI. It is the plumbing: getting the right data to the right tool, and proving each step did what it claimed.