Key takeaways
- A custom dashboard tracks six assets (BTC, ETH, SOL, NVDA, TTE, CW8), pulls prices, computes technical indicators and pushes three written briefs a day to Telegram, at 8:50am, 1:00pm and 5:30pm Paris time.
- The writing runs on Claude Haiku 4.5, an Anthropic model released on October 15, 2025 and priced at $1 per million input tokens and $5 per million output tokens: a daily brief costs a fraction of a cent.
- NVIDIA researchers have made the same argument since June 2025: for an agent's narrow, repetitive tasks, small language models are good enough and far cheaper than large ones.
- The underlying stack (collect, compute, write, schedule) carries over to any SMB as is: a morning sales brief from the CRM, a midday stock check, a weekly cash summary.
The problem: six assets, ten tabs, decisions made in the heat of the moment
The starting point is personal. Seriously tracking six assets across three families, three cryptocurrencies, two stocks and one ETF, meant several browser tabs permanently open, technical indicators recomputed by hand, and constant back-and-forth between platforms that do not speak the same language.
The real cost was not access to the data: prices are public and free. The cost was assembly. Every check meant mentally rebuilding a full picture from fragments, and that rebuilding usually happened at the worst possible time, in reaction to a sharp move. A decision made in the heat of the moment, staring at a red chart, is rarely a good one.
The need fit in one sentence: have the information arrive assembled, written up, at times decided in advance, instead of chasing it in a hurry.
What was built: a pipeline that ends with writing
The dashboard does three things, in order. It collects prices for the six assets. It computes technical indicators from those prices, with conventional, deterministic, verifiable code. Then comes the step that changes the nature of the tool: a language model turns those figures into three written briefs a day, sent to Telegram at the times that matter, 8:50am before European markets open, 1:00pm at midday, 5:30pm after the close.
Around that core, several modules round out the daily use: Telegram alerts with three severity levels for moves that cannot wait for the next brief, a scenario simulator, a tracker for recurring investments, a market events calendar and a trading journal that keeps a record of decisions.
The architectural point that matters: the LLM (Large Language Model) invents nothing and computes nothing. It writes from figures established upstream. Reliability comes from the pipeline; the model brings readability. That separation is what keeps the setup lean and predictable.
Why is a small model enough?
Because the task is narrow, repetitive and fully framed: same assets, same indicators, same output format, three times a day. That task profile does not call for a frontier model.
The model in use is Claude Haiku 4.5, priced by Anthropic at $1 per million input tokens and $5 per million output tokens, and positioned as the fastest in the range. Anthropic itself lists financial analysis and monitoring large numbers of data streams among its typical use cases. At that rate, a brief of a few hundred tokens costs a fraction of a cent; the budget for a tool like this sits in the build, not in the running costs.
The choice lines up with a broader finding. In a research paper submitted on June 2, 2025 and revised in September 2026, NVIDIA researchers argue that small language models are sufficiently capable, better suited and markedly more economical than large ones for the specialized, repetitive tasks of agentic systems. A dashboard's daily brief is exactly that kind of task.
The table below sums up the options available to anyone who wants to follow several assets, and where each one stops.
| Option | What it covers | Where it stops |
|---|---|---|
| Broker app | Live prices, simple threshold alerts | One asset at a time, no written overview |
| Charting platform with alerts | Technical indicators, conditional alerts | Each alert is a raw signal you interpret yourself, the moment it fires |
| Custom dashboard with an LLM | Your assets, your indicators, three written briefs a day on a fixed schedule | Needs building and maintaining; overkill for occasional tracking |
And inside an SMB? Same stack, different data
The head of a small or mid-sized company is in exactly the position of someone holding six assets. The numbers exist: in the CRM, in the ERP, in the bank account, in the inventory system. They are simply scattered across tools nobody opens spontaneously in the morning. The dashboard often already exists; it is the habit of opening it that never took hold.
The transposition is direct. A sales brief at 8:50am: deals that moved yesterday, overdue follow-ups, leads left without an answer, pulled from the CRM and written up in ten lines. A production or inventory check at midday. A cash summary on Friday. The mechanics are identical to the market dashboard: collect, compute, write, deliver where teams already read, whether that is Telegram, Slack or email.
The difference with a classic BI (Business Intelligence) report comes down to one reversal: the reader no longer goes looking for the information; it arrives written, at the chosen time. A dashboard nobody opens is useless; a ten-line brief that lands in your pocket gets read.
When this setup is a bad idea
Custom software is not always the answer, and this case is no exception. If you look at your numbers once a week and they fit in a spreadsheet, keep the spreadsheet. If your BI tool or your CRM already sends scheduled reports your teams actually read, do not add a layer on top.
Two more limits deserve stating plainly. If the underlying data is unreliable, a well-written brief will only dress up wrong numbers; the real work is upstream, in data entry and structure. And a language model producing automated investment recommendations without supervision is not a good idea either: in this setup, the tool describes, the human decides.
If your numbers live in three tools nobody opens, the check is simple: for one week, count how many times you went looking for a piece of information you would rather have received. To talk it through, book a 30-minute call, and see how I build this kind of tool on the custom web app page. At minimum you will leave with an honest opinion, including when the right answer is the scheduled report in your current tool.
Frequently asked questions
What exactly does the LLM do in this kind of dashboard?
The language model writes; it does not compute. Prices are collected and technical indicators calculated upstream by conventional, verifiable code; the model turns those already-established figures into a short, readable brief. It makes no decisions and gives no investment advice.
How much does running an automated LLM brief cost?
With a model like Claude Haiku 4.5, priced by Anthropic at $1 per million input tokens and $5 per million output tokens, a daily brief of a few hundred tokens costs a fraction of a cent. Most of the budget goes into building the pipeline. At Mathieu Haye, a Sprint starts from €400 excl. VAT; a full dashboard is quoted per project.
Do you need a frontier model to analyze business data?
Not for narrow, repetitive summaries. NVIDIA researchers have argued since June 2025 that small language models are powerful enough and far more economical for the specialized tasks of an agent. A large model earns its keep on open-ended, exploratory or multi-document analysis.