Skip to content
ayan://bhardwaj (home)

    Type a code and press Enter, or pick a key.

    DES <GO> · Independent build · Finance

    YouBank

    The AI deal desk that learns how you work.

    It does the work and learns how you do it: a keyboard-driven research terminal on SEC data, a Studio that builds the model and the deck while you watch, and an email agent that has to earn the right to send on its own.

    Open youbank.dev (opens in a new tab)Code : GitHub (opens in a new tab)

    Free while in beta.

    youbank.dev
    YouBank landing page: a terminal mock with a watchlist and quarterly revenue bars beside counters for tools, startups, workflows and learning loops
    terminal functions
    35+
    tools: 230 AI workflows, 98 calculators
    328
    recall@5, up from 0.85 with cross-encoder reranking
    0.97
    of drafts go out unedited
    90%+
    unit tests
    1,000+
    styles, 7 dark and 7 light
    14

    ENGINE <GO>

    Autonomy is earned, not switched on.

    Every draft you send is a label. YouBank measures how much you changed it: word-level edit distance over the longer draft. A draft counts as good when that ratio is 0.05 or less and no critical field moved: a number, an amount, a percentage, a link, an email address, a weekday or a month.

    ENGINE · earned autonomy · kind: “scheduling reply”
    50%60%70%80%90%100%autopilot bar
    Curve: the engine's belief about how often this kind of draft goes out unchanged (Beta posterior, uniform prior). Shaded: 90% interval. Dot: the value it is 90% sure the rate is above. It must clear the dashed bar.

    No decisions yet

    Decisions
    0
    Counted good
    0
    Estimate
    50%
    90% interval
    5%–95%
    Would need edits
    ≤ 90%
    1. Each sent draft is a label.

    Autopilot is offered for a kind of email only after at least 20 decisions, and only when the engine is 90% confident that at most 10% of those emails would need edits. The engine keeps a Beta posterior per kind, with its own Beta CDF and quantile, Lanczos log-gamma and a seeded generator. The demo below runs the same rule.

    1. Loop 1Earned autonomyYour edits decide, per kind of email, when the agent may send alone.
    2. Loop 2Outreach experimentsThompson sampling over opening angles and send times, rewarded by replies and penalised by opt-outs.
    3. Loop 3Lessons from editsWhat you change in a draft becomes guidance for the next one.

    FUNC <GO>

    A terminal where the models are written down.

    More than 35 functions, Bloomberg style. Company functions take a ticker, market ones don't, and some take plain words. Each screen can get an AI read in three bullets. A sample, with the methods behind them:

    Terminal functions, how to type them, and the methods behind them
    CodeWhat it doesMethodsTyped as
    GPPrice chart with a price coneGARCH(1,1)SNOW GP
    BETABeta, four waysOLS, Blume, Welch (2022)SNOW BETA
    RISKValue at risk and expected shortfallEWMA, GARCH, Cornish-Fisher, Kupiec backtestSNOW RISK
    WACCCost of capitalCAPM, implied-rating spread, Monte Carlo, tornadoSNOW WACC
    IRATImplied credit ratingAltman Z'', Ohlson O, Merton distance to defaultSNOW IRAT
    QUALEarnings qualityBeneish M, Piotroski F, Sloan accrualsSNOW QUAL
    FCSTForecasts with honest intervalsM4 combination, conformal intervalsSNOW FCST
    EEOdds of an earnings beatBeta-BinomialSNOW EE
    GCTreasury curve and recession oddsNelson-Siegel, NY Fed recession probitGC
    ECOMacro dashboardDamped trend, Sahm ruleECO
    EQSPlain-English screener over every SEC filerWords in, a screen outEQS profitable software companies growing over 20%
    PORTPortfolio riskLedoit-Wolf shrinkagePORT AAPL 40 MSFT 30 KO 30
    BTXBacktest grader, A to FDeflated Sharpe, PBO by CSCV, Harvey-Liu haircutBTX
    FACTFactor LabFama-French 2x3 on a point-in-time top-1,000 universeFACT
    LEARNLearns which functions you knowBayesian knowledge tracing: hints fade and new functions appear as you learnLEARN

    Research tools only: YouBank is not investment advice and never places orders.

    STUDIO <GO>

    Watch the model build itself.

    Studio is a live workbook and deck. The agent builds a DCF, LBO, comps or pitch book from SEC data cell by cell while you watch: cells glow as they are written, the agent's cursor is visible, and the view follows it across sheets.

    Slides are linked to the model, so changing WACC moves every page. Each run ends with an audit health strip and an undo for the whole run. Deep mode runs up to 40 steps plus a review pass, and it exports to .xlsx and .pptx or works inside Excel and PowerPoint through an add-in.

    STUDIO · Example Co. · DCF.xlsx + Pitch.pptxRun complete
    Illustrative discounted cash flow model, millions of dollars, WACC 9.5%
    $MFY26EFY27EFY28EFY29EFY30E
    Revenue1,2401,4881,7411,9852,203
    Growth24.0%20.0%17.0%14.0%11.0%
    EBIT margin8.0%11.0%14.0%16.0%18.0%
    EBIT99164244318397
    Taxes(21)(34)(51)(67)(83)
    NOPAT78129193251313
    D&A5060707988
    Capex(62)(74)(87)(99)(110)
    Change in NWC(24)(25)(25)(24)(22)
    Unlevered FCF4290150207269
    Discount factor0.9130.8340.7620.6960.635
    PV of FCF3875114144171

    Slide 4 · Valuation summary

    $3.25B

    Enterprise value at 9.5% WACC, 3% terminal growth. Terminal value is 83% of it.

    Slide 5 · WACC sensitivity

    • 8%$4.35B
    • 9%$3.56B
    • 10%$2.99B
    • 11%$2.57B
    • 12%$2.25B

    Slide 6 · Free cash flow build

    Free cash flow rises from 42 to 269 million dollars.

    ▲ Audit 60 cells written · 0 errors · 0 hard-codes in the forecast · deck ties to model

    A toy version written for this page. Example Co. and its inputs are made up; the real Studio pulls its numbers from SEC filings.

    AGENT <GO>

    An email agent that asks before it guesses.

    The email agent reads your mailbox, triages it and drafts in your voice. When it cannot answer something it asks you one question and keeps the answer in a playbook. It also runs outbound sequences from a directory of 18,000+ startups and keeps a pipeline from lead to won.

    Only one function in the codebase sends mail, and it re-reads the live thread first. Anything the engine is unsure about is held in a review queue with its reasons.

    1. New mail
    2. Sync
    3. Resolve thread
    4. Triage
    5. Draft
    6. Consider
    7. Held with reasons, or scheduled
    8. Re-read live thread
    9. sendDraft()the only code that sends
    10. learnFromSend
    Relationships · your queue
    The agent's morning queue: a question card, high and medium confidence drafts, and an autopilot countdown

    NEWS <GO>

    A newsroom that reads the filings first.

    Every ten minutes the Newsroom reads about 90 free sources (publisher headlines, EDGAR filing signals, regulators and wires, and a radar per sector) and turns them into stories for each desk. Its public front page is open to anyone.

    NEWS · wire format every 10 min

    1. EDGAR8-K · Item 1.03· Restructuring

      A bankruptcy filing goes to the top of the restructuring desk

    2. EDGARSC 13D· Activism

      An activist crosses 5%, with the stake, the date and the letter attached

    3. EDGARMerger proxy· M&A

      A merger proxy is parsed into the Precedent Engine, every term with its quote

    4. EDGARS-1· ECM

      A new IPO registration is read for the numbers that matter

    Example cards showing the kinds of signals it reads, not live stories.

    youbank.dev/news
    YouBank's public Newsroom front page with serif headlines and story cards

    BUILD <GO>

    How it is built.

    One person, one TypeScript monorepo with Python and Rust where they earn their place. The README keeps a dated research log and the docs folder holds the persona research, decisions and specs.

    App
    Next.js 16 App Router, React 19, TypeScript, Tailwind v4 with generated theme tokens
    Data
    Neon Postgres through Drizzle, with 29 hand-written idempotent SQL migrations
    Auth
    Better Auth, self-hosted: Google and email/password
    AI
    OpenAI Responses and Anthropic Messages APIs with zod structured output; every model call logged with its token cost
    Sources
    SEC EDGAR submissions, XBRL company facts and full-text search; Financial Modeling Prep
    Jobs
    Vercel Cron for the morning sync and agent run, Neon Functions triggers every 20 minutes for autopilot and news
    Mail
    IMAP/SMTP or the Gmail API; mailbox credentials encrypted with AES-256-GCM
    ML
    PyTorch services on Modal: a GraphSAGE model for M&A link prediction, and hybrid retrieval with cross-encoder reranking that lifted recall@5 from 0.85 to 0.97
    ml/
    Python cores for LBO, odds and edge models, with tests, and an optional GPU retraining job on Modal
    Quant
    Models written from scratch in TypeScript: GARCH volatility, VaR with backtests, Monte Carlo valuation, a Treasury yield curve, Altman, Ohlson and Merton credit risk
    Tests
    1,000+ unit tests
    desktop/
    Tauri 2 app in Rust: tray, youbank:// deep links, a global Quick ask hotkey, an opt-in local agent
    Office
    Office.js add-in for Excel and PowerPoint
    • Next.js
    • React
    • TypeScript
    • Tailwind v4
    • Neon Postgres
    • Drizzle
    • Better Auth
    • OpenAI + Anthropic
    • zod
    • Python
    • PyTorch
    • Modal
    • Rust
    • Tauri 2
    • Office.js
    • Vercel

    SCRN <GO>

    Screens.

    From the public site, plus two signed-in views on a phone that a visitor can't reach without an account.

    The real engine demo
    The real engine demo on youbank.dev after 22 decisions: the posterior curve sits past the 90% autopilot bar and the badge reads Earned autopilot
    A trading comps table with EV to revenue, Rule of 40, sparklines and a scatter plot
    Studio feature cards and the build checklist: building the model from SEC data, writing cells, tying out the deck
    The gallery of fourteen styles, seven dark and seven light
    The signed-in terminal on a phone, listing function codes
    The signed-in Studio on a phone