A portfolio management platform for a family office

A desk that runs a few hundred positions across many brokerage accounts, on margin, with a pipeline of names it is researching and a weekly rhythm of decisions, needs more than a brokerage portal. It needs one book that reconciles to the custodian every morning, a way to look at every position against the same yardstick, a record of why each name is held and who argued for it, and a feed of what moved and why that is already written by the time anyone sits down. For years that lived in a workbook that was rebuilt every week. This platform replaced it, and it has grown with the desk for the better part of a year of daily use.

The day has a shape. Overnight, the custodian's encrypted feed lands and is reconciled into tax lots; prices, fundamentals, statements and news are refreshed; the monitors run and the attention memory decides what is new. By the time the desk opens the Daily Review, the Compare heat map has scored every position, the fundamentals monitor has fired on what crossed a limit, and the morning email has already said which fifteen names deserve a look and why. During the day the Review Board carries the arguments: analysts vote on the pipeline and the portfolio, the portfolio manager converts a call into a trade with one leg per account, and the trade queue is the only door to the custodian. After the close, the movers email reads the whole market and the snapshot reads the book.

What makes it different is that it argues with itself. A red cell is red only when the number is strictly past the limit the monitor fires at, so the table and the alerts can never disagree. Every measure carries its definition and its last four quarters, so a reader can check a number before acting on it. A thesis has exactly one writer and a version history. A topic report tiers every listed name in a space and says, for each one, what it rests on and what would break it. And nothing automated can trade: the AI reads, scores, drafts and flags, and a person decides.

The platform replaced a spreadsheet-based workflow for a multi-account book with a nightly broker feed, live pricing, a research pipeline, a daily review system and a set of automated briefings. Every screen in this demo is the real template rendered over an invented book: the tickers are real public companies so the charts and fundamentals are genuine, but the accounts, lots, notes, votes, theses and people are synthetic. The AI-written passages (earnings previews and reads, desk briefs, KPI cards, mover notes) are composed for the demo from those public figures and a short profile of each company; nothing is generated live, and none of it is a recommendation.

Some of the parts worth a look:

• The Compare heat map scores every position against the same limits the fundamentals monitor fires at, and an attention memory tracks whether each red cell is new, worse, chronic or already seen, so the morning email leads with what changed rather than with what is merely red. Right-click a cell to mark it seen; click a reading to see its last four quarters.
• The Review Board keeps the pipeline and the portfolio on one table. Analysts vote; a trade is filed only when the portfolio manager converts a call, never by an AI and never by agreement alone. The report pane prints a one-pager per stock with evidence chips captured from the figures that made the case.
• The research layer gives every idea a provenance, every stock a versioned thesis with a single writer, and every theme a generated report that discovers, verifies and tiers the listed names in a space, with odds, differentiation and both cases written for every one.
• The operations half runs on a separate worker process: the broker feed, overnight enrichment, cache warming, twenty-odd scheduled emails, and backups with an immutable off-site copy that has been proven by rebuilding the whole platform from it.

Python and Flask with SQLAlchemy on PostgreSQL; Jinja templates with Tailwind, Alpine.js for the reactive parts and Chart.js and Lightweight Charts for the charts. One web service and one worker process, so a long import or a model sweep can never take a page down; a scheduler with a dozen nightly jobs, each wall-clock bounded and watched by a heartbeat.

• Data in. The custodian's PGP-encrypted SFTP feed every morning, with a self-healing backfill for a late day; yfinance for prices, fundamentals, statements and corporate actions, with the corporate-action cliffs re-synced automatically; SEC EDGAR for 13F filings; Finnhub and RSS for news; a free-text-proof ticker universe.
• Models where reading is the job. Haiku for classification and screens, Sonnet for judgement, Perplexity for search, each metered and cached where a cache can actually hit; a reviewer that reads a stock's own dossier before it is shown the desk's thesis; prompts that forbid recommendations where a recommendation would be a second opinion.
• Rules that keep the numbers honest. Red means strictly past the limit; a missing value sorts last in both directions; a failed fetch is never recorded as data; a percent and a fraction never meet in one variable; every rollup is a prefix test over one tree.
• Five thousand five hundred tests, checked by mutation: each defect a test guards against is re-introduced on purpose to prove the test can fail, and the ones that could not are rewritten. Templates are rendered under strict undefined and their inline scripts parsed by Node in the suite.
• Operations. A deploy on every push through a serialised safe-deploy that waits out a running import; nightly backups of every database to two object stores, one of them immutable for thirty days; an in-app assistant that reads the book through the app's own calculations, never by recomputing them in SQL.

The code itself is private. If you would like to walk through it, get in touch.