What you are looking at

The desk has five fictional clients, each in a different country and currency: an endowment in the US, a family office in the UK, a foundation in Switzerland, a pension scheme in Germany and a mutual insurer in Japan. Each holds one portfolio of real instruments: US equities, cryptocurrencies, and in one case a listed bitcoin option. Every portfolio is valued in its client's own currency and its value at risk is checked against a limit, continuously.

Two services run it. A feeds service publishes the reference data, the simulated FX and the desk's standing instructions to Kafka. The desk is the Calseda application: it reads those, reads the New York Fed's overnight rates straight from its API, connects to the exchanges and lookup services on demand, and runs the calculations. The console you are in is the desk's built-in view of itself.

Five ideas

Everything on the desk is built from five things.

Discovery and live inputs

Each worker gets its data in two phases, and the difference matters. In discovery, it asks for data step by step, and each answer decides the next question. In live inputs, it names the values its formula uses, and its formula runs again whenever one of them changes.

1. Discovery, step 1 The portfolio: which holdings, which client.
2. Step 2 The client and, for each stock, its instrument record and trading venue.
3. Step 3 The countries of the client and of each venue, which give the currencies.
4. Live inputs The price of each holding, and only the FX rates the conversion needs.

If discovery data changes (a holding is added, a client moves), the worker is rebuilt and starts discovery again, so it now watches the new holding's price. If a live input changes (a price ticks), only the formula runs. Open any worker in the console to see both phases drawn as a data flow.

The tasks

The desk runs nine kinds of task here, and four more for derivatives, described in the next section. The diagram shows how these nine connect: whose results each one reads, and which ones a browser page asks for.

Derivatives

Four kinds of worker turn free market data into the figures a derivatives desk checks every day. They use Deribit, the largest exchange for bitcoin and ether options, plus the US Treasury's yield curve and the New York Fed's overnight dollar rate. Their results are on the Options page. A few ideas first, since everything below builds on them.

A short primer

An option gives its holder the right, but not the duty, to buy (a call) or sell (a put) something at a fixed price, the strike, on a fixed date, the expiry. A bitcoin call with a strike of $90,000 pays off only if bitcoin ends above $90,000; a put with a strike of $70,000 pays off only if it ends below $70,000. Deribit writes these as BTC-27NOV26-90000-C and BTC-27NOV26-70000-P.

Implied volatility. An option pays off most when prices move a lot, so its price tells you how much the market expects the price to swing before the expiry. Turned into a yearly percentage, that is the option's implied volatility: 40% means moves of roughly 40% over a year, or about 11% over a month. It is how traders compare options, much as a yield is how they compare bonds.

The smile. Options on the same date but with different strikes imply different volatilities. Drawn against the strike, they usually form a U-shaped curve: the market pays extra for options that only pay off after a large move. In crypto, protection against falls is usually dearer than bets on rises, so the curve leans to one side.

Futures and the perpetual. A future is a contract to buy or sell at a price agreed now for delivery on a later date. Its price is the market's best estimate of the coin's price on that date, the forward, and options on that date are priced against it. A perpetual is a future that never expires; instead, its buyers and sellers pay each other a small funding rate every 8 hours, which keeps its price close to the coin's.

Free data limits this to crypto: live prices for options on shares are sold by the exchanges, and there is no free source for currency options, interest-rate options or credit. The same workers would run on a customer's own data for any of them.

An LLM in the loop

The filing analysis worker is an ordinary Calseda worker that calls a language model. It reads text no deterministic code could, and fills in a fixed form: whether the filing is material, and each event's type, direction, severity and a supporting quote. Nothing else it writes is used.

• Contained. The model only reads real SEC filings: nothing a visitor types reaches it. Its answer is forced into the form, checked again in code, and a quote is shown only if it really appears in the filing.
• Deterministic after it. A fixed table turns the events into a price shock, and ordinary workers apply it to live positions and VaR limits. Those numbers move with the market; the model is not asked again.
• Read once, shared. Each filing is analysed once. The result is stored on Kafka, reused after restarts, and read by every worker and page that needs it, so model calls grow with filings, not with users.
• Tunable without code. The prompt is data, published with the reference data. A new prompt makes the framework recreate the analysis workers, which re-read their filings with it.

Use it from your AI agent

The desk is also an MCP server, so an AI agent can use it directly: see what the desk computes, read live results, ask why a portfolio is over its limit, and run a pre-trade check, a what-if valuation or a filing's impact, privately, as these pages do. The tools and their inputs are generated from the desk's configuration, with no code written for agents. Connect an agent shows how to connect Claude, Cursor, VS Code, Gemini or Codex, and lists the tools. Then ask, for example, "Which portfolio is over its VaR limit, and why?" or "What would buying 1 BTC do to P001's risk?"

Why it stays efficient

A desk like this could easily do far more work than it needs to: stream every instrument, recompute on every message, republish every result. Calseda avoids that in five ways, all visible in this demo.

Feeds open only on demand

An exchange subscription exists only while some worker needs that instrument, and is released 30 seconds after the last one stops. Sell a holding and its price feed closes; nobody pays for data nobody uses, and free-tier rate limits go a long way.

Only the inputs a calculation uses

Each worker names exactly what its formula reads, so a price tick re-runs only the workers that hold that instrument, and an FX tick only those whose conversion needs that rate.

Unchanged values do no work

A republished value that hasn't changed triggers nothing: no recalculation, no rebuild, no new result. A result that comes out the same as last time isn't published again.

Latest values, not a queue

When prices move faster than a calculation runs, it doesn't work through a backlog of stale updates. It runs on the latest value of every input and skips the intermediate ones, so results stay current under load instead of falling behind. This is conflation rather than dropping work at random: no input is ever older than its latest value.

Rebuild only when the shape changes

Only a change to discovery data, such as a new holding, rebuilds a worker. Price moves never do.

What it means for you

A five-minute tour

1. Open Valuation and pick P002. Its data flow shows the three discovery steps and the live inputs: two stock prices, bitcoin and two FX rates, because the client is in the UK and the holdings are priced in dollars.
2. Open Value at risk. One or more portfolios are usually above their limit; the result shows each position's contribution and the volatility used.
3. Build a portfolio on What-if valuation and watch it value live. Change the client and the currency changes with it.
4. On Pre-trade check, try buying bitcoin for a stock-heavy portfolio and see the VaR move against the limit.
5. Open Filing review to see recent filings by held companies, matched to their legal entities and parents.
6. On Filing analysis, pick a company's earnings 8-K: see what the model found, the shock the rule table gives it, and what that means for each holder against its VaR limit.
7. Under Entities, every data type lists its source, and every key its current value.

Simplifications

• US stock prices only move while US markets are open; at other times valuations move with crypto and FX.
• FX is simulated around the ECB's daily reference rates.
• The VaR model is deliberately simple: parametric, per position, with no diversification.
• The filing shock table is illustrative, and earnings are judged against the prior year, not analysts' expectations.
• It is a demo on one machine, not a sizing or performance benchmark.