A decision-intelligence layer, built for horticulture
Twenty-four AI features, and every one exists because of a Harvora prediction. The findings are deterministic — the AI puts them into words. Grounded, auditable, and always yours to confirm.
No AI output touches a number it didn't retrieve
Every AI-generated report, briefing, insight and answer runs the same five-step discipline. Hallucinated figures aren't rare — they're architecturally impossible.
Every relevant prediction, observation and dollar figure — including actual costs synced read-only from Xero — is retrieved from your blocks' database first.
Assembled into a structured context bundle — the AI only ever sees verified figures.
The model synthesises and narrates. It never computes a number on its own.
Output is parsed against a strict schema before a grower ever sees it.
If validation fails, a deterministic structured display ships instead. No malformed AI output, ever.
Five detectors sweep every block, every night
At 2:30 am the nightly scan runs deterministic detectors over each block's prediction history. Only genuine findings surface — deduplicated against alerts you've already seen, narrated in plain language by morning. It also plans: the AI Season Planner reschedules frost, irrigation, harvest and scouting events on every forecast update; your 5 am Daily Briefing ranks the day's decisions by expected value; and when you open an event, a pre-decision checklist spells out what to verify before you act. Weekly, pre-season and debrief reports round out the record.
The forecast midpoint has shifted beyond the block's own rolling variability.
Agronomist-reviewed rules across surfaces — e.g. sustained disease risk while yield softens.
Today's value sits far outside the same block at the same point in past seasons.
Cold hours, disease-favourable days or water deficit building without any single-day alert firing.
A prediction lands outside the confidence band the block's own season model expects.
The urgent calls, delivered once — while there is time to act
Critical and watch findings are bundled into one coordinated SMS and email digest. Harvora suppresses repeats, carries escalation forward, respects quiet hours and only breaks the hold when waiting would close the action window.
Every digest starts with a real critical or watch item from the same analysis layer that drives Today's Priority.
Fingerprint deduplication prevents reworded repeats, while a watch escalating to critical supersedes the pending alert instead of double-paging.
Alerts are bundled and rate-limited per recipient, with quiet hours holding delivery unless waiting would miss the action window.
Both channels carry the same digest at the same moment, from one delivery clock and one source of truth.
Ask Harvora — an agent that knows your growing operation
Ask anything about your operation and get an answer grounded in your blocks, your history, your numbers — right down to actual costs synced read-only from Xero. Under the hood it's a bounded agentic loop — powerful enough to run what-if scenarios, disciplined enough to never act without you.
Operation bundle v3 reconciles every relevant prediction and calendar item by provenance. Prediction identity and validity window travel with the evidence; current, superseded, expired or unresolved states are resolved before an answer is trusted, and claim-level grounding shows exactly which parts of the response are supported.
Drafts, answers and proposes — never acts without you.
Every critical answer carries typed prediction identity, model version and a validity window. Related calendar evidence is reconciled as current, superseded, expired or unresolved before the agent responds.
A bounded agent loop over read tools, a scenario runner and tiered action tools — hard budgets, one repair retry, deterministic fallback.
Hybrid full-text + vector search over your own records, rank-fused, isolated to your orchard in SQL. Citations shown.
High-stakes questions are answered three times independently and you get the consensus; if the runs disagree, that disagreement is flagged to you rather than hidden.
Hands-free in the orchard — questions spoken, answers streamed back sentence by sentence.
External agents can read and propose through a consent-gated endpoint. They can never execute a write.
A reliability layer under all 24 features
An end-to-end adversarial audit hardened every reliability edge of the AI layer — the part that keeps it trustworthy under load, bad input and provider failure. Each item below shipped with tests, and the model-facing ones with a passing evaluation.
Facts the model reads from a chat, an image or a retrieved document never auto-write to your profile — they queue as confirm/reject suggestions, so a hidden instruction can't quietly poison every future answer.
A truncated or malformed generation says so and degrades cleanly, instead of stopping mid-sentence or masquerading as an outage.
High-stakes answers are generated multiple times and the response you see is the one the runs agree on — judged by what the answers mean, not how they're worded.
A rolling per-orchard cost budget covers even the agent stream and block ask; context and tool results are capped, with an explicit “trimmed” note rather than a silent drop.
Provider retries preserve useful coverage, partial recovery stays visible as recovering, and optional unconfigured services are kept distinct from genuine degradation.
Intelligence that compounds, block by block
Every season you grow with Harvora, four learned models deepen for each block. Each carries an explicit insufficient-data status — nothing is fabricated on a young block. What they learn flows back through the whole engine: forecast corrections, expectation bands, risk thresholds and briefings, all tuned to how each block actually behaves — predictions that keep getting sharper, built on a track record no competitor can replicate without watching your blocks through the same seasons.
Observed temperature and humidity are compared to the regional forecast under each synoptic pattern — clear nights, advection fog, nor'west flow — and the block's systematic deviation becomes a correction the frost model applies.
Which lagged weather conditions precede elevated Psa pressure on this block — learned per block, because aspect and shelter change leaf wetness independently of the region.
How this block's yield responds to weather at each phenological stage, calibrated against your confirmed harvests. It needs two seasons of confirmed yields — the regional model carries it until then.
The actual soil-moisture evolution of this block under rain and evapotranspiration — correcting FAO-56 for soil properties and drainage the map can't see.
The AI drafts.
You decide.
Harvora's AI suggests, schedules, and generates — but nothing writes to your operation without your explicit confirmation. Human judgment stays exactly where it belongs.
Work with us ↗