What we shipped, what's next.

The first cut was FinTerm: a capabilities dashboard, a frosted-glass theme, and dark mode set as the default inside the same day. An AI assistant went in the following morning with market summaries attached, and a unified integration layer followed on the 13th alongside input validation and response headers. The model was the thing on display, which turned out to be the wrong promise to make.

The rename landed across app, marketing and assets in one release, on top of 22 shared UI primitives with a gallery to check them against. Core screens were migrated onto the primitives rather than reskinned, and the shell picked up mobile navigation. The indigo accent that shipped here as the identity did not survive: the terminal is amber on near-black today, and the glass material of these first weeks is now banned in product outright.

An assistant API with persistent history, deploys gated by health checks so a bad build cannot take the site down, a staging environment with its own safety rails and backups, and a status page with real health endpoints behind it. Unglamorous, and the reason the next four months could run at a release every two days.

Two releases in one day put the platform in front of people who had not signed up. The homepage was rebuilt around a live insider-flow timeline drawn from real SEC filings, beside public ticker pages, a public screener with a fast cold start, an embeddable ticker widget, and financial statements with hardened parsing. Living Analyst Notes shipped with them as the first flagship AI surface.

Real data from more than twenty government and public sources: SEC EDGAR, FRED, ClinicalTrials.gov, USPTO and the rest. Underneath it went production error tracking, structured logging, a deep health check across core systems, code splitting for the initial load, and server-side preferences so theme, sidebar and layout follow you between devices.

Freshness badges spread across data surfaces over four releases, then collapsed into a single shared component that replaced every older variant. Insider data consolidated onto one canonical source read by every surface, patent data onto another, and the portfolio learned to flag a brokerage connection older than 24 hours. This is the run where the rule was set: a data surface states how fresh it is, in the same words, everywhere it appears.

Insider coverage widened to the whole S&P 500 with a heatmap widget, alongside broader 10-Q, 10-K and 8-K coverage. Alerts arrived with email delivery and one-click unsubscribe, CSV export came to the screener and watchlists, and semantic search went over the filings corpus.

The assistant was renamed Brief and given your linked brokerage tickers, so it could answer what do I own. Holdings started flowing into correlation, stress test and VaR without being re-entered anywhere, grouped by account on the dashboard, with a one-tap refresh and per-tab sync indicators.

The redesigned terminal shell became the default, bringing a full-fidelity stock-detail rebuild, an Insider Flow overhaul with per-row filing links, peer comparison on a nightly refresh, and real-time market data on top of the existing API. A Sectors page added rotation, breadth and dispersion, and a workflows hub added saved research. Then the assistant, which had been scattered across pages, consolidated into a single persistent, page-aware sidebar.

Add to coverage appeared on every stock page, tier-gated, with a fetch fast enough to run while you waited, and a newly added name populated immediately across the core sources. Fundamentals moved to deriving from filings data rather than a per-name vendor call, which covered nearly the full universe and refreshed far faster. The Financials tab went from populated for a few hundred companies to populated for almost every covered one.

Public endpoints for reference, key stats, events, related companies, history and the screener came first. Self-serve key management, a usage dashboard and public documentation followed, with Pro Research credit packs as one-time top-ups and Morning Brief available as a standalone subscription. Natural-language screening arrived in the same release, turning a plain-English prompt into screener criteria with evidence citations on the answer.

Dashboard v2 shipped as the default with a live signal feed and a data-freshness badge on every panel, then reorganized around four trader moments rather than around whatever happened to be available. Filing search landed beside it, and the assistant learned to search and cite material-event filings.

The score was rebuilt for a sharper, less noisy signal, two weeks after a three-factor Conviction Score v2. Stock charts gained intraday data and roughly twenty years of daily history, earnings detail reached forty-plus quarters with real revenue-surprise figures, and the financials table was rebuilt around a readable key-ratios strip.

Multi-factor regression on a portfolio, tax-lot accounting in FIFO, LIFO and average cost with wash-sale detection, a mean-variance and risk-parity optimizer with a rebalancer, and backtesting with walk-forward, Monte Carlo and parameter sweeps. Exposure intelligence shipped on its own route with live satellite and vessel layers over curated disruption regions, and the Options Lab hub opened with volatility context and a payoff backdrop.

A data-time badge, an as-of note on anything the AI wrote, and a per-value source trail you can open on a figure. Business Owner Mode put an owner-earnings view on stock detail, a revenue-segment explorer showed mix over time and per-segment year-over-year, and custom dashboards, common-size statements and custom metrics arrived together.

Lobbying, retail and Reddit chatter, search trends and corporate-jet tracking joined the feed roster. A unified notifications inbox collected web push, options-flow alerts and thesis alerts into one place instead of a surface per feature, and the account page was rebuilt with billing and usage embedded in it.

IV rank, skew, term structure and volatility-risk premium in a Volatility Lab; a market-wide Greeks scanner; an earnings model showing the pre-earnings implied-volatility ramp and the crush after; a strategy optimizer that ranks structures from a plain-language thesis; and a backtester pricing at real end-of-day implied volatility. The same release opened programmatic access, including an MCP server, so an assistant outside Convexity can call its tools directly.

A verdict on a stock with the evidence shown on every axis, a history timeline, a watchlist feed, a free preview and a shareable public receipt, over a red-flag score computed across filing text. Any cited figure in an assistant answer became clickable through to the exact source passage, and answers began defaulting to tables with one-click copy into Excel or CSV.

Portfolio Overview was rebuilt on one unified book of record, leading with value, day P&L, unrealized gain and concentration, with an analytics row for blended fundamentals, performance attribution, liquidity and dividend income. A cockpit followed carrying a benchmark overlay, asset-class allocation and a P&L toggle. Beside it came an earnings-call reader with speaker and section parsing, where ask this call is grounded in the transcript you have open and cites it back.

The Situation Room was rebuilt as an immersive globe with fly-to navigation, a live geo-briefing and an ops basemap, over an ontology linking world events to the companies and mechanisms they affect. Three days later Atlas Situations began clustering correlated cross-domain anomalies into one tabular situation with an evidence trail and an activation chain, and Mimir why-paths traced the shortest receipted route between any two entities, replayable as of a past date.

Saved charts, a formula editor, a catalog of forty-plus FRED macro series and company fundamentals, native candlestick and heatmap renderers, and embeddable chart cards. A publication pass a month later added annotation styles, axis chrome, a legend, a one-click IPO line, drafts that persist and callouts that point at a specific value, with AI generation writing in the same styling and taking a refine pass.

One canonical page per company with nine sourced sections, reachable from exposure rows and from every stock page. A filings desk added a sentence-level redline between two versions of a filing, extracted financial tables, a topic timeline, and investor presentations pulled out of 8-K exhibits. Every significant mover started carrying the news that best explains it, with a same-sector fallback and an explicit flag when a volume-backed move has no headline behind it at all.

Structured theses had shipped four days earlier: your case, catalysts, risks, valuation and invalidation triggers, saved to your account and monitored as the world moves. The ledger made them keep score. Every stance change is recorded against what the market did next, a nightly scorer runs the book, and a track-record page reads your calibration over time. A view that grades itself is a different object from a note.

Plain-English criteria wired straight into the filter rail, above a hero scatter, a range rail and a command surface. The Earnings lens was rebuilt as Next, Last and History, composed on the server so the three panels cannot disagree with each other, and the Convexity Score gained a why-this-score popover with sector-relative A+ to F factor grades behind it.

A nearest-neighbor panel across eleven dimensions that returns real companies, the historical analogs behind them and validated outcome statistics, and that discloses its match quality instead of handing back the least-bad five. The company graph became walkable in the same release: reverse lookup, in-canvas peer expansion, draggable cards and a what-changed read, with knowledge time recorded so any view replays as it stood on a past date.

Alerts began delivering into Slack and Discord channels, written for each rather than dumped as a payload, and scoped to the person who owns the rule so a shared endpoint cannot leak another account. Everything the names on a watchlist publish arrived in a single feed instead of five surfaces, and the exposure graph learned people and institutions: officers, auditors drawn from audit-engagement filings, fund domiciles, portfolio tilts and disclosed sale intents.

Every tool the assistant had was retrieval. It could report any figure the backend had already computed and could not compute one nobody had anticipated. It now writes analysis and runs it in a throwaway sandbox, against datasets fetched server-side through the same canonical queries the rest of the terminal reads. The model never carries the data.

A Vault on the terminal to search your own documents, then run a review grid that answers a whole column of questions across a stack of them at once. Word, PowerPoint and Excel joined PDF and CSV on the document on-ramp three days later, with slide numbers surviving the trip so a retrieved passage can say where in the deck it came from.

Earnings call history went from three years to ten, filled a slice at a time so one deep backfill could not crowd out the feeds sharing the same upstream budget. Every fiscal period on the statements table began naming the filing that reported it, and saying so plainly when no filing resolves. Operational KPIs, the subscribers and ARR and deliveries companies report on the call and never file in a statement, were read out into a quotable per-quarter series with the sentence they came from attached.

The read-only tool server had shipped tools for months and could be reached only by minting an API key by hand. It moved behind a standard authorization flow, so a client connects with your account. Then a legibility pass: the lit amber left rail came off every row, card and panel, content moved into one column against the bezel, and every chart primitive learned to answer a pointer with the value under it, from one shared implementation rather than three hand-written crosshairs.

The free tier had been closed by an accident of routing rather than a decision: every sign-up button on the site led to a waitlist form. Signing up now creates a free account and signs you straight in. A one-question chooser at first sign-in (research a company, monitor a watchlist, connect a portfolio, find an idea) starts you on the page that does it and keeps the next step on the dashboard, and a paid plan picked on the pricing page survives sign-up to the upgrade page.

Every lookup the assistant makes now appears live while it works, so an answer arrives with the trail that produced it. Each answer can be copied, retried or taken deeper on the company it is about, and the composer learned the terminal: @ completes a ticker, a slash picks the kind of answer, and selected text in an answer offers explain, source, compare and copy.

Saved screens moved from the browser to the account, so a screen saved at one desk opens at another. Thesis coverage now starts from the book you already hold, largest uncovered position first, with a review action and a count of the live theses you reviewed in the last seven days. The Research tab on a company opens on your own history with it, and standing work got an off switch: alerts, scheduled prompts and living reports pause and resume from one page.

One company, one timeline: 8-K and periodic filings, insider and Form 144 filings, court dockets and analyst rating changes, each scored for materiality and traced to its source. The part that matters most is what it refuses to imply. A lens whose feed does not carry the company says so instead of reading as a quiet stretch, because a clean timeline is only worth something when you know what was watching. Two more lenses joined in September: guidance revisions, and the metrics a company starts or stops giving on its calls.

Describe an edge from insider buying, estimate revisions, price moves or federal contract awards, and test it only against what was public on each past date, beside a same-date control. Then arm it, and it keeps a live record separate from the backtest, so what a rule did after you believed it can be read apart from what it did in hindsight. Below thirty observations it declines to estimate at all.

The War Room now leads with whatever matters at that point in the reporting cycle. Before a print it sets what the options market implies against how the stock moved on past reports; after it, the numbers against consensus and the call. Every section it cannot cover is labelled rather than left blank, and the rebuild caught an earnings surprise that had been displayed a hundred times too small.

SEC filings listed by the second EDGAR accepted them, within minutes, with the full text of every press release a company furnished with an 8-K or 6-K beside them. Each 8-K row names the items it reports, and the reader searches and steps through a release and shows it in the layout the filer gave it, tables intact. Developer and Quant plans can have each filing pushed by webhook, and press releases filed with the SEC became searchable in full text the same week.

Decks that companies post on their own investor sites are read slide by slide rather than linked: they open in place on the stock page, document search finds the slide that says what you searched for, and each one has a reader page that sets out its slides in order. Collection widened through the fortnight, from hosted investor feeds to ordinary investor pages. Earnings calls gained the start time the company stated on its own webcast platform, which calendar feeds mostly give only as a day.

A prompt became an object. Save one from the composer, reuse it with a slash, fill its placeholders before it sends, and point a schedule at it so that editing the prompt changes what runs next. Within two days a schedule could wait for the document instead of the clock, running when the transcript or quarterly filing of a watched company arrives, or when a chosen type of document lands.

A Models lens on the stock page values the company seven ways: two discounted cash flow models, three multiples, excess returns and a dividend discount model. It shows where each lands against the price and how far apart they sit, and says why a method could not run, a bank skipping the cash-flow and enterprise-value methods or a company that pays no dividend skipping the last, rather than leaving a gap to be read as a result.

What a company said it would do became a record you can follow. The Guidance Screen shows which companies guided a line item and to what, beside the sentence each figure came from, and the Guidance Ledger marks each call as raised, cut, reiterated, narrowed or widened, naming the pairs it cannot compare instead of calling them unchanged. Company State and the Peer Matrix shipped beside them, and by mid-September KPI extraction had widened from the 150 largest companies to the 2,000 largest while guidance history doubled to eight quarters.

Earnings call transcripts reached past the 500 largest companies toward the 6,000 largest, and quarterly and annual reports were indexed across the same broad universe as 8-Ks. The Documents page gained the releases, reports, supplements and guidance updates companies post on their own sites, then proxy statements and the annual and current reports of foreign private issuers. Then it learned to list each document once, however many tickers or addresses carry it, and removed what had been collected from sites that belonged to another company.