An autonomous pre-market screen for coffee-futures buyers. Every morning before the bell — today's best buy zone and the optimal time window for the front month plus the next seven contracts of Arabica (KC) and Robusta (RC). Sixteen contracts. One refresh. Honest confidence.
GRINDER is the workbench the desk sits at before the open. Click REFRESH and the agent pulls the KC and RC forward chains, fits a 35th-quantile ATR alpha against 120 sessions of back-adjusted continuous data, intersects each target with a floor-trader pivot guardrail, and reports the densest 60-minute window in which the session low has historically occurred. Each of the sixteen rows carries a target price, a buy band, a bucketed confidence tag (HIGH · MED · LOW — never a fake percentage), and the time window with its observed hit rate. Where Robusta data is proxied from KC, the row says so explicitly — a DATA-PROXIED badge surfaces every inferred input. The model fits in ~6–9 seconds, runs from a desktop browser, and saves an audit-trail snapshot every refresh so any decision can be replayed against the inputs it was made with. Sibling workflow to WARREN — same Lualdi design language, different domain.
Coffee procurement desks live between two clocks. The roaster's blend calendar runs in weeks. The futures market runs in minutes. Between them sits the question that determines the year's coffee P&L: at what price, on what day, in what window, should we bid? Most desks answer that question with a chart, a hunch, and a phone call to a broker. GRINDER answers it with a fitted alpha, a pivot guardrail, an empirical time-of-day distribution, and an audit trail — every morning, in the time it takes to drink the first espresso.
Front month plus the next seven for both Arabica (KC) and Robusta (RC). Sixteen rows, sixteen targets, sixteen bands, sixteen time windows. The desk no longer triages which contracts to look at — the workflow looks at all of them, every refresh, before the bell.
Every target is published with a ±0.15·ATR band and an intersection against the [S2, P] pivot guardrail. The output is a range the desk can actually bid into, with a floor that respects yesterday's structural support — not a single point estimate that pretends to a precision the model doesn't have.
HIGH, MED, LOW. Three booleans: quiet volatility, small overnight gap, target inside the pivot guardrail. The desk reads the bucket and knows what it's looking at. We do not display a calibrated probability because we have not earned the right to: with 120 sessions of fit data, a bucket is honest and a percentage is theatre.
An empirical histogram of when the daily low has occurred on 1-minute bars over the last 5–7 sessions, smoothed with a triangular kernel. The densest 60-minute window is reported alongside the observed hit rate. The desk knows when, not just where, to be ready.
A click on any row opens the dossier drawer — the model inputs, the pivot levels, the ATR, the alpha quantile, the session sample for the time-window histogram, and a Haiku-written desk note that reads the row in plain English. Every refresh is snapshotted; any decision can be replayed against the inputs it was made with.
Robusta deferred-history is not yet freely available. RC's alpha is proxied from KC continuous, the time window is proxied from KC=F intraday, and the RC-front ATR is a one-day true range. Every substitution surfaces as a DATA-PROXIED badge on the row. The desk sees exactly which inputs are first-class and which are inferred — no silent fallbacks.
GRINDER is a deterministic engine on top of a small, well-behaved statistical model. Every input is observable. Every step is replayable. The published row is the consequence of the recipe, not the recipe.
KC forward chain from Investing.com's server-rendered table (eight contracts). KC daily OHLC history from yfinance — KC=F for the continuous front, KCxxNN.NYB tickers for each deferred. KC intraday 1-minute bars for the time-window histogram. RC front from Investing.com. All cached with a TTL appropriate to the refresh cadence; cache misses surface explicitly.
For each contract: α = q35((open − low) / ATR14 | gap_sign) over 120 sessions of back-adjusted continuous data. Target = open_proxy − α · ATR14_contract. Buy zone = [target − 0.15·ATR, target + 0.15·ATR] intersected with [S2, P] — floor-trader pivots from yesterday's H/L/C. The pivot guardrail caps the band so the model never recommends a bid below structural support.
Three booleans → bucket: ATR percentile low, |gap|/ATR < 0.4, target inside [S2, P]. 3/3 = HIGH, 2/3 = MED, ≤1/3 = LOW. Time window = empirical histogram of the daily-low timestamp on 1-minute bars over the last 5–7 sessions, smoothed with a triangular kernel. Densest 60-minute window reported with observed hit rate.
GRINDER runs in a desktop browser. One screen, one refresh button, three tabs. The dense-grid canvas is a Lualdi design language signature — terminal-grade, instrument-grade, built for an operator who already knows what they're looking for.
Sixteen rows — eight KC, eight RC — keyed by contract month. Per row: target price, buy band, confidence bucket, time window, hit rate, and the badges that flag proxied inputs. Sort by confidence, by month, by contract. Click any row to open the dossier. The whole snapshot is timestamped with UTC, age, and the model version that produced it.
Click a row, the drawer slides in. Inside: every model input — yesterday's H/L/C, the pivot levels, the ATR(14), the gap, the alpha quantile, the time-window histogram with its session sample — plus a Haiku-written desk note (Claude Haiku 4.5) that reads the row in plain English. Every snapshot is saved to disk; any decision can be replayed against the inputs it was made with.
Speciality and commercial roasters with active hedging programmes against forward sales. GRINDER's row is the morning input the procurement manager reads before any bid is placed — sized against the blend coverage forecast already produced by ORIGINS.
Green-coffee importers and proprietary trading desks running a futures-overlay against physical inventory. GRINDER's deferred coverage matters here — the desk hedges across the curve, not just the front, and reads sixteen rows rather than one.
Single-family offices and macro allocators with coffee exposure as part of a broader softs book. GRINDER's audit trail and bucketed confidence make it auditable for an investment committee that asks "why did we bid there, on that day, in that window?"
Private beta access by introduction · Private deployments via the engagement desk · Response within two business days