Daffodil

AI-Powered Vegetation Line Prediction for Indoor & Controlled-Environment Farms

Market Focus: New York State (CEA / Vertical Farming Corridor)

New York has become one of the densest hubs for Controlled Environment Agriculture (CEA) in the U.S. — large vertical farms upstate (e.g., the Hudson Valley, Buffalo, and Newburgh corridor) and rooftop/warehouse farms in NYC. These operations grow indoors under LED arrays, hydroponic/aeroponic racks, and tightly controlled climate zones — but they still make planting, thinning, and harvest-line decisions largely by eye or fixed schedules, not by real-time plant growth data.

Daffodil's core idea: use computer vision + environmental sensor fusion to predict the "vegetation line" — the growth-front boundary showing which rows/trays are ready to advance to the next stage (germination → vegetative → harvest) — days before it's visually obvious to a human.

Target customer: indoor/vertical farm operators, greenhouse growers, and CEA facility managers across NY (5–500k sq ft facilities).

2. Core Features

🌟 Flagship Feature: Predictive Growth-Line Forecasting

Instead of just showing current plant health (what most agtech dashboards do), Daffodil forecasts where the vegetation line will be in 3, 7, and 14 days, using:

• Overhead/rack-mounted multispectral cameras (NDVI-style indoor imaging)
• Time-series growth-rate modeling per tray/row, calibrated per crop variety
• Environmental inputs (light hours, temp, humidity, CO2, nutrient dosing logs)

Output: a visual "growth-front map" of the facility with a countdown to harvest-readiness per zone, so labor and logistics can be scheduled in advance rather than reactively.

Supporting Features

3. Implementation Plan

Phase 1 — Hardware Integration (Weeks 1–4)

• Install fixed or rail-mounted multispectral/RGB cameras above grow racks
• Connect to existing climate/sensor systems (Priva, Argus, TrolMaster, etc. via API/Modbus)
• Edge device (NVIDIA Jetson-class) per zone for local image preprocessing

Phase 2 — Model Calibration (Weeks 3–8, overlaps Phase 1)

• Baseline data collection per crop variety grown (leafy greens, herbs, microgreens, strawberries, etc.)
• Train/fine-tune growth-curve prediction models per crop using facility's own historical + live data
• Human-in-the-loop correction: growers tag "ready" vs "not ready" for the first 2–3 cycles to refine accuracy

Phase 3 — Dashboard & Alerts Rollout (Weeks 6–10)

• Web dashboard (facility map, growth-line visualization, forecasts)
• Mobile app for floor staff (zone-level alerts, checklist, harvest queue)
• Integration with labor scheduling tools (e.g., When I Work, or CSV export)

Phase 4 — Optimization Loop (Ongoing)

• Monthly model retraining as more harvest cycles complete
• Expansion to yield forecasting and buyer-commitment tools
• Optional: integration with NY distributor/wholesale marketplaces for pre-harvest sales

Team & Timeline

• MVP for single facility: ~10 weeks
• Team needed: 1 CV/ML engineer, 1 backend/infra engineer, 1 frontend engineer, 1 agronomist (part-time, crop calibration), 1 installer/technician

4. Pricing Breakdown

Pricing is structured by facility size (sq ft of active grow space), since that drives camera/sensor count and compute load.

One-time costs:

• Hardware install (cameras, edge devices, cabling): $3,000–$15,000 depending on facility size (can be leased instead: +$150–$600/mo)
• Onboarding & model calibration: $2,500 flat fee (waived on annual contracts)

Add-ons:

• Extra crop variety model training: $400/variety (one-time)
• SMS alerting (beyond email/push): $50/mo
• NY buyer-marketplace integration: $200/mo

Discounts:

• Annual prepay: 15% off
• NY-based CEA co-ops / multi-farm consortiums: 10% additional volume discount

5. Go-to-Market Note

New York offers relevant angles worth leaning on in sales conversations: state CEA grant programs (via Empire State Development and NYSERDA energy-efficiency incentives for indoor ag), and the concentration of vertical farms in the Hudson Valley/Capital Region makes for an efficient regional sales and installation loop before expanding to other states with strong CEA activity (New Jersey, Ohio, Michigan).

Core principle: separate the edge (per-facility, low-latency) from the cloud (shared, heavy compute) from day one. Most agtech startups couple these and pay for it at scale-up.

1. Edge Layer (per facility)

• Jetson-class device per zone runs a lightweight inference model (quantized, e.g. TensorRT) for real-time image capture + basic anomaly detection — works even if internet drops.
• Sensor/camera data buffered locally (SQLite/MQTT broker), synced to cloud on interval, not continuously streamed. Keeps bandwidth and cloud ingestion costs flat as you add facilities.
• Why this matters for scale: compute cost scales with facility count, not centrally — you're not paying for one giant inference cluster serving every camera in real time.

2. Ingestion & Data Layer

• Ingestion: Kafka (or managed equivalent — MSK/Confluent) as the backbone. Every facility publishes to the same topic schema regardless of hardware vendor (Priva, Argus, TrolMaster) — normalize at the edge, not in the warehouse.
• Storage split:
  • Raw images/spectral data → object storage (S3), lifecycle-tiered (hot 30 days, cold after)
  • Time-series sensor data → a TSDB (TimescaleDB or InfluxDB) — this is what your growth curves are built on
  • Structured/relational (facilities, zones, crop cycles, users, billing) → Postgres
• This 3-way split is the single biggest scaling decision: don't put time-series in Postgres, don't put images in a database at all.

3. ML Layer (the actual product)

• Feature store (Feast or homegrown) sitting between raw data and models — this is what lets you retrain per-crop, per-facility models without duplicating pipelines.
• Per-crop-variety models, not one global model — leafy greens grow nothing like strawberries. Train a base model, fine-tune per facility with their calibration data (your Phase 2 human-in-the-loop labels).
• Training: batch, offline, scheduled (weekly retrain as cycles complete) — not real-time. Real-time is only inference at the edge.
• Model registry + versioning (MLflow) from day one, even for one facility — you'll thank yourself at facility #10 when you need to know which model version made which prediction.

4. Application Layer

• API layer (FastAPI/Node) — stateless, horizontally scalable, sits in front of Postgres + TSDB
• Multi-tenancy from the start: every table keyed by facility_id, even with one customer. Retrofitting multi-tenancy later is the most common rebuild I see in this space.
• Dashboard (React) + mobile app both hit the same API — no separate backends.

5. Scaling Path

6. MLOps / Reliability

• CI/CD for models, not just code — every retrain gets validated against a holdout set before it's pushed to any facility's edge device.
• Drift monitoring: alert when a facility's live growth curves diverge from the model's training distribution (common when a grower switches crop variety or lighting recipe).

The single most important early decision is the edge/cloud split with per-crop model architecture — it's what lets you onboard facility #2 without re-engineering, and it's the part most competitors in this space get wrong by centralizing everything.

Calibration data is what turns a generic "plant growth" model into one that actually knows this facility's strawberries under this facility's lights. Here's what's actually in it:

1. Paired image + label data

For each tray/row, at each imaging timestamp:

• The raw multispectral/RGB image
• A human-assigned ground-truth label: "not ready" / "approaching" / "ready to harvest" / "past peak"
• Ideally a continuous score too (e.g., estimated days-to-harvest), not just a category — this is what lets you train a regression model instead of just a classifier

This comes from growers tagging trays on the floor for the first 2–3 crop cycles — either through the mobile app (quick tap: "this row is ready") or a technician doing a manual pass with a tablet alongside the cameras.

2. Environmental context at time of imaging

Each labeled image gets tagged with the conditions that produced that plant state:

• Cumulative light hours (DLI — daily light integral)
• Temperature/humidity history for that zone since planting
• CO2 levels
• Nutrient dosing log (EC/pH, feed schedule)
• Days since germination/transplant

This matters because "ready" doesn't mean the same thing at day 21 under 16-hour light as it does at day 21 under 12-hour light. Without this context, the model just memorizes "day 21 = ready" and breaks the moment the grower changes their light recipe.

3. Crop variety metadata

• Variety/cultivar name (e.g., "Rex Butterhead" vs "Salanova Green")
• Seed lot / supplier, if the facility tracks it (genetic variation affects growth rate)
• Planting density per tray

4. Outcome data (the feedback loop)

• Actual harvest date and yield weight per tray
• Any post-harvest quality flags (bolting, tip burn, mold) — these become negative examples that help the model learn what not to call "ready"

Concrete example row

facility_id: NY-004
zone_id: B-12
crop_variety: Butterhead Lettuce (Rex)
image_timestamp: 2026-07-03T09:00
image_ref: s3://.../B-12_20260703_0900.tif
days_since_transplant: 18
DLI_cumulative: 14.2 mol/m²/day
temp_avg_7d: 68°F
humidity_avg_7d: 62%
grower_label: "approaching" 
grower_est_days_to_harvest: 4
actual_harvest_date: 2026-07-07
actual_yield_g: 142
quality_flag: none

That row is one training example. You need a few hundred per crop variety before the model is more useful than a fixed schedule — which is why Phase 2 calibration takes 2–3 full cycles rather than being a one-time setup step.

Here's a fuller breakdown of each tier and what lives inside it — building on the architecture from before:

Tier 1: Edge Layer (per facility)

Purpose: low-latency capture and first-pass inference, works offline.

Tier 2: Ingestion & Data Layer

Purpose: get normalized data from many facilities into the right storage, cheaply and reliably.

Tier 3: ML Layer

Purpose: turn raw + calibration data into predictions.

Tier 4: Application Layer

Purpose: expose predictions to humans and other systems.

Tier 5: MLOps / Reliability (cross-cutting)

Purpose: keep models trustworthy as facilities and crop variety count grows.