A perennial-crop health monitoring platform.
Final Course Project (TCC) submitted to the Institute of Technology and Leadership (INTELI) to obtain a bachelor's degree in Computer Engineering.
Authors: Eduardo França Porto · Marcos Vinicyus Rosa Teixeira Advisor: Prof. Rodrigo Nicola Team G112 · T21 · São Paulo · 2025/2A
Project completed. This README follows the structure of INTELI's Final Course Project template (Entrepreneurial track, NBR 14724).
CropTrack is a B2B computer-vision platform that detects coffee leaf diseases and pests early, from images or video, without proprietary hardware. The work pivoted from a from-scratch CNN classifier (fragile in the field, 60–70%) to a fine-tuned YOLOv8 object detector, which answers where and how much disease is present and reached 90.3% mAP@50 (precision 85.8%, recall 85.4%) on four coffee classes. Around the model, a field platform was built (React + Flask): georeferenced plot map, Manager → Collector → Manager workflow, asynchronous video analysis with notification, and an agronomic dashboard. The market was sized (TAM R$ 1.8 bi / SAM R$ 114 M / SOM R$ 1.2 M) and a per-hectare SaaS model defined (R$ 5–10/ha/month). Validation was qualitative and measured (NPS +50, n = 8 agronomists), with a real distribution channel (Casa da Roça, PA) and a field pilot (Jaguaré, ES). The problem was externally validated by Y Combinator's 2026 Request for Startups. The project is pre-pilot: field ground-truth validation is the next step.
Keywords: computer vision; object detection; coffee; precision agriculture; software as a service.
Global food security is, this century, fundamentally a productivity problem. The United Nations projects the world population will grow from 8.2 billion (2025) to 9.8 billion by 2050 — an additional 1.6 billion people, most of them in emerging economies whose diets are simultaneously becoming more protein- and coffee-intensive. The Food and Agriculture Organization (FAO) estimates that meeting this demand requires global agricultural output to rise 60–70% by 2050. Crucially, there is no longer abundant arable land to expand into: of the projected production growth, the FAO attributes 87% to productivity gains (more per hectare), 7% to higher cropping intensity and only 6% to area expansion. The strategic conclusion is direct — the future of food depends on producing more with what already exists, and the first step is to stop losing what is already planted.
The losses are large and concrete. In Brazil, pests and diseases cost agriculture an estimated R$ 55–60 billion per year (Agrolink/EMBRAPA), and productivity drops of up to 40% are documented depending on the crop and region. In coffee specifically, diseases such as leaf rust, leaf miner and brown eye spot (cercospora) cause losses of roughly R$ 4.5 billion per year (≈ US$ 900 million; EMBRAPA Café/MAPA). Brazil is the world's largest producer and exporter of coffee (~34% of global output), with ~330,000 coffee properties and 1.9 million hectares in production — which means each percentage point of productivity lost in Brazil reverberates through the entire global supply chain.
Despite the availability and falling cost of technology, the field remains largely analog. Disease detection is still done by manual scouting: an agronomist walks a fraction of the plot, by eye, on a fortnightly basis. This is expensive, slow and sub-sampled, and it pushes the producer to spray fungicide on a fixed calendar — out of fear, not out of data — with no institutional memory of what happened in each plot across harvests. Adoption is the bottleneck, not the technology: ~78% of Brazilian coffee producers are family operations with low digitalization, so any solution must be simple and hardware-free to be adopted at all.
A unique technological window now makes the solution viable: open-source computer vision models (YOLOv8) trainable with modest data, near-zero cloud inference cost, expanding rural 4G coverage, affordable cameras and drones, and regulatory pressure for traceability (EU Deforestation Regulation). External validation came in 2026, when Y Combinator named "AI for Low-Pesticide Agriculture" its #1 Request for Startups — a near-verbatim description of the CropTrack thesis.
General objective. Create and validate a computer-vision solution for early detection of diseases and pests in coffee leaves, and develop a business plan for its introduction to the market.
Specific objectives.
- Develop the detection MVP (a fine-tuned YOLOv8 model) with field-usable confidence (a target of ~85%+ on the chosen metric);
- Build the field platform around the model: georeferenced plot map, image and video analysis, role-based workflow and an agronomic dashboard;
- Size the addressable market (TAM/SAM/SOM) with primary sources and define a defensible revenue model;
- Validate the problem and the solution qualitatively (interviews and usability/NPS) and secure a real distribution channel and a field pilot;
- Map the competitive landscape and articulate the sustainable competitive advantage.
Document structure. Section 2 develops the solution: it states the market premises and the problem/solution/value hypotheses (2.1), sizes the market and profiles the customer (2.2), analyzes competitors and differentiators (2.3), describes the technological solution and the MVP results (2.4), presents the business plan (2.5) and reports the market validation (2.6). Section 3 concludes, stating whether the objectives were achieved, the honest limitations and the future projections for the venture.
The project was structured around three explicit, falsifiable hypotheses, in the lean-startup tradition — a problem hypothesis, a solution hypothesis and a value hypothesis — each tied to a way of being validated.
Problem hypothesis. Coffee producers detect disease too late. Detection today relies on manual scouting, which is sub-sampled by construction (the agronomist sees a fraction of the plot) and reactive. The economic consequence is twofold: first, producers over-apply fungicide preventively, paying for chemicals they may not need; second, when a focus of disease such as leaf rust is finally noticed, the loss has already compounded. On a 200-hectare farm, late rust detection can mean R$ 160,000–500,000 of lost production in a single harvest — far more than any monitoring technology would cost. There is also no per-plot memory: each harvest restarts from zero, with no historical basis to guide management.
Solution hypothesis. AI detection over a cheap camera, georeferenced, delivers an early and localized diagnosis. By running an object-detection model (YOLOv8) over a photo or video captured with a phone or drone, the platform returns not a single label but bounding boxes that answer where the disease is in the plot and how much of it there is. Tied to a georeferenced plot map, this transforms a diagnosis into a management decision (targeted application, re-inspection), without any proprietary hardware — addressing the adoption barrier directly.
Value hypothesis. Producers accept a per-hectare SaaS because the saving dwarfs the price. The agribusiness already reasons in cost-per-hectare (inputs, labor, harvest), so charging per hectare is the natural unit. The anchoring is strong: avoiding a single unnecessary fungicide application on 200 hectares saves R$ 30,000–50,000, while the platform costs on the order of R$ 12,000/year for the same farm — an ROI in the range of 4× to 12×, i.e., the tool pays for itself many times over with one avoided spray.
These hypotheses were not left theoretical: Section 2.6 reports how each was probed through interviews, a usability test with measured NPS, and the securing of a real channel and pilot.
The market was sized top-down, starting from the total perennial-crop opportunity in Brazil and funneling by technological eligibility and early-stage conversion capacity. Pricing in this sizing uses the Premium reference of R$ 10/ha/month (the price accepted by producers who already pay for aircraft/drone services — see 2.5).
TAM — Total Addressable Market. Brazil has roughly 15 million hectares of perennial crops with potential use (coffee, citrus, eucalyptus, cocoa, fruit). At R$ 10/ha/month over twelve months, the total addressable market is:
15,000,000 ha × R$ 10/ha/month × 12 = R$ 1.8 billion/year
Restricting to coffee in production alone (1.9 million hectares, EMBRAPA/MAPA 2024) gives an addressable coffee-only figure of ~R$ 228 million/year — the immediate, most defensible beachhead within the larger perennial TAM.
SAM — Serviceable Addressable Market. Not every hectare is reachable with the current product. The eligibility criteria are: farm area above 50 ha (so per-hectare SaaS is viable), field connectivity (rural 4G or fiber), a responsible agronomist, and the capacity to pay without public subsidy. Applying these filters yields roughly 950,000 eligible hectares of coffee plus around 100 medium/large cooperatives:
950,000 ha × R$ 10/ha/month × 12 = R$ 114 million/year
(plus ~100 cooperatives as a white-label channel layer)
SAM ≈ R$ 114 million/year
SOM — Serviceable Obtainable Market. The share realistically capturable in the first 12–18 months with the current team and resources. The base scenario is 50 clients with an average of 200 ha:
50 clients × 200 ha × R$ 10/ha/month × 12 = R$ 1.2 million/year
This implies an implicit conversion of 0.25% of the ~20,000 eligible farms — a conservative figure for a pre-seed stage, consistent with Brazilian AgTech benchmarks (e.g., Strider reached 50–80 clients in its first 18 months).
TAM R$ 1.8 bi/year (15M ha of perennials)
└ SAM R$ 114 M/year (coffee > 50 ha + cooperatives)
└ SOM R$ 1.2 M/year (50 clients · 12–18 months)
Customer segmentation and profiling. Three priority segments were defined:
| Segment | Decision profile | Pain | Sales cycle |
|---|---|---|---|
| Large farms (> 50 ha) | Centralized in the manager agronomist | Loss from late detection; over-spraying | 2–4 weeks |
| Coffee cooperatives | Innovation/technical committee | Need to serve hundreds of associates and prove sustainable management | 4–8 weeks |
| Agronomist consultants (RTs) | The consultant himself | Differentiate the service; cover the whole plot | Per service |
The cooperative is strategically the highest-leverage channel — a single contract reaches hundreds of producers and enables a white-label offer; the agronomist RT, in turn, is both a user and an acquisition channel for end clients.
The competitive landscape was mapped on two axes — diagnosis precision (low → high) and adoption cost (low → high, including price, hardware dependency and learning curve). CropTrack occupies the ideal quadrant: high precision at low adoption cost, with no proprietary hardware.
| Solution | AI diagnosis | Plot map | No proprietary HW | Coffee BR | Est. ticket |
|---|---|---|---|---|---|
| CropTrack | ✅ YOLOv8 | ✅ | ✅ | ✅ | R$ 5–10/ha/month |
| Aegro | ❌ | ❌ | ✅ | Partial | R$ 529+/month |
| Solinftec | ❌ | Partial | ✅ HW | ❌ | R$ 2,000+ |
| Strider | ❌ (manual) | ✅ | ✅ | Partial | R$ 200–500 |
| Agronow | Partial (NDVI) | Partial | ✅ | ❌ | Negotiated |
| Plantix | ✅ (phone) | ❌ | ✅ | ❌ | Free |
| Taranis | ✅ (drone) | ✅ | Drone | ❌ | US$ 5–15/ha/yr |
| Agrio | ✅ (phone) | ❌ | ✅ | ❌ | Free |
Read individually: Aegro is a farm-management ERP with no disease vision (a potential distribution partner, not a competitor); Solinftec does prescriptive spraying but is hardware-dependent and does not diagnose disease visually nor serve coffee; Strider digitizes integrated pest management but the diagnosis remains 100% manual; Agronow uses satellite NDVI whose resolution (meters) cannot detect disease on an individual leaf; Plantix and Agrio do phone-photo diagnosis but without georeferencing, plot map or history (and Plantix's user is not the B2B payer); Taranis is the closest in capability (drone AI) but costs 5–10× more and has no structured Brazilian operation.
Sustainable advantage (moat). Anyone can train a YOLO model — what cannot be copied is (i) proprietary Brazilian field data that improves with every harvest; (ii) the switching cost created by the per-plot history locked in the platform; (iii) the cooperative white-label channel, which produces a network effect; and (iv) coffee as a beachhead, the most demanding validation environment, which credentials expansion into the other perennials.
Development methodology. The project ran on an agile (Scrum) cadence across ten two-week sprints — five technical and five business. The technical phase (Sprints 1–5) delivered the model and the platform; the business phase (Sprints 6–10) delivered market analysis, go-to-market materials, impact analysis and the final consolidation.
The technical journey (a deliberate pivot). The first approach (Sprints 1–3) was a from-scratch classification stack: custom CNNs (CustomCNN, and variants inspired by ResNet/EfficientNet) over a large balanced coffee-leaf dataset. On the public test set the best model reached very high accuracy, but two limits emerged: it fell to 60–70% in realistic field conditions (the classifiers learned global colour shortcuts and failed to generalize), and, conceptually, classification answers the wrong question — is the leaf sick? — without locating or quantifying the focus in the plot. The decision (Sprint 5) was to pivot from classification to object detection, fine-tuning YOLOv8. Detection returns where and how much, with a far smaller data requirement than training from scratch, and reached the field-usable confidence target. Recognizing the ceiling of the first attempt and migrating deliberately is itself one of the project's main engineering results.
MVP — modules and features. Around the model, a full field platform was built (React frontend, Flask backend):
- AI engine: YOLOv8 detection over photo and video, with annotated bounding boxes and multiple specialized detectors;
- Georeferenced plot map (Leaflet + satellite imagery): the manager draws the plot polygon and releases collection spots;
- Role-based field workflow (Manager → Collector → Manager): release, field collection, and validation, persisted across sessions;
- Asynchronous video analysis: long videos are processed in the background; the user can navigate away and is notified on completion, with the annotated result persisted and available on the manager's dashboard;
- Agronomic dashboard: health distribution, alerts, weather, and a per-plot activity timeline ("Field Life").
Test results (final model). The production detector is a YOLOv8n fine-tuned for four coffee classes, with ≈ 3.01 million parameters and ~6 MB of weights — small enough to run on CPU and, in the future, on the collector's device:
| Metric | Value |
|---|---|
| mAP@50 | 90.3% |
| mAP@50-95 | 61.5% |
| Precision | 85.8% |
| Recall | 85.4% |
The 90.3% mAP@50 comfortably exceeds the ~85% confidence threshold set as a condition for a real management decision, and contrasts with a generalist multi-disease detector (33% mAP@50 over 29 classes) — evidence that focusing on coffee was the correct technical choice.
Production architecture (designed for scale). Although the deliverable is a monolithic prototype, the topology was designed to evolve in the cloud: an inference microservice separated from the API, fed by a queue (SQS) with stateless, auto-scaling workers (scale-to-zero); object storage (S3) for media and artifacts; a relational database (Postgres + PostGIS) for plots and analyses; and notifications (SNS). Because YOLOv8n is small and CPU-friendly, the analysis is inexpensive — the estimated marginal cost is under US$ 0.01 per hectare analyzed — and a pilot can operate for ~US$ 150–270/month (or ~US$ 20–60/month in a serverless footprint).
The business model is per-hectare SaaS, priced by value, with three tiers:
| Plan | Price | Audience | Value |
|---|---|---|---|
| Entry | R$ 5/ha/month | Consultants and smaller farms (skeptics) | Quick diagnosis, low commitment |
| Premium | R$ 10/ha/month | Producers already using aircraft/drone | Scales with the property |
| Cooperative | Negotiated (min. R$ 5,000/month) | Multi-tenant for associates | White-label, reports |
Price anchoring. The price is anchored on the alternatives the customer already pays: a manual technical visit (R$ 300–600), a full fungicide application on 200 ha (R$ 30,000–50,000, 3–6×/year), monthly agronomic consulting (R$ 1,500–4,000). Against these, R$ 5–10/ha/month is a fraction of the cost — and a small fraction of the loss it prevents.
Break-even and unit economics. Monthly operating cost (infrastructure plus one dedicated person) is on the order of R$ 4,200–8,000. The product becomes self-sustaining with 25–30 clients on the Premium plan — a target achievable in the first year. The ROI for the client (4–12×) sustains the willingness to pay, and the value anchor (one avoided spray) keeps the price well below the customer's perceived value ceiling.
Financial projection (labeled, base scenario). With documented assumptions (CAC R$ 500–1,500; annual churn 15–25%, a B2B-SMB SaaS benchmark), the base projection is:
| Metric | 12 months | 24 months | 36 months |
|---|---|---|---|
| Active clients | 15 | 45 | 100 |
| MRR | R$ 8k | R$ 24k | R$ 55k |
| Break-even | — | Month 18–22 | — |
These are projections, explicitly labeled as such; the venture is pre-revenue and the churn/CAC figures are benchmarks, not measured values.
Channels. Four channels, prioritized by segment: cooperatives (white-label, the highest-leverage); the family network Casa da Roça (9 agro stores in southeastern Pará, an immediate low-CAC channel); agronomist RTs (the tool differentiates their service); and sector events plus digital content for inbound.
Validation methodology. Validation was deliberately empirical rather than theoretical: (i) pain interviews with producers and agronomists to confirm the problem hypothesis; (ii) a usability test of the application with measured NPS; and (iii) the active securing of a distribution channel and a design partner for the field pilot, to confirm the path to real adoption.
Market validation results.
- NPS +50 (n = 8 agronomists) — a measured, not projected, metric: of the eight respondents, the distribution skews to promoters, confirming receptivity.
- Pain confirmed: the eight interviews consistently reported late detection, manual scouting and preventive spending "out of fear" — corroborating the problem hypothesis.
- Distribution channel in hand: the Casa da Roça network — 9 agro stores in southeastern Pará, a family business — is willing to sign a letter of intent as a channel, giving direct access to hundreds of producers with no initial CAC.
- Design partner / pilot: the family coffee farm in Jaguaré (ES) is the site of the first pilot — the path to the first real field ground truth for the model.
- External validation: Y Combinator's 2026 Request for Startups placed "AI for Low-Pesticide Agriculture" as its #1 priority, validating the problem at a global level of authority.
Taken together, these results move the project beyond a theoretical exercise: the problem is validated externally and qualitatively, and both the channel and the field are already accessible — the conditions for converting the first contracts.
The objectives set out for the project were achieved. On the technical side, the deliberate pivot from classification to YOLOv8 detection took the model from a fragile proof of concept to a detector with 90.3% mAP@50, embedded in a usable, hardware-free field platform (map, role-based workflow, asynchronous video, dashboard). On the business side, the problem was validated externally (Y Combinator) and qualitatively (NPS +50, interviews), the market was sized with primary sources (TAM R$ 1.8 bi / SAM R$ 114 M / SOM R$ 1.2 M), a defensible per-hectare revenue model was defined, and a real distribution channel (Casa da Roça) and field pilot (Jaguaré) were secured. More than a model, CropTrack is a product coherent with a real problem, built with transparency about its stage.
Limitations (stated honestly). The model's metrics come from the training validation set, not from field-collected ground truth — performance under field shade, dust and background still needs to be measured. The venture is pre-pilot, with no paying revenue, so financial figures are projections labeled as such. The dataset carries imbalance and a laboratory → field gap, and visually similar diseases (cercospora × phoma) remain a confusion risk.
Future projections. The immediate next step is the field pilot in Jaguaré to generate real ground truth and convert the Casa da Roça letter of intent into the first contracts. Beyond that: continuous re-training with the proprietary field data that constitutes the moat; edge inference (the model running on the collector's phone) for offline use where rural connectivity is poor; the cloud production architecture described in 2.4; and expansion from coffee into citrus, eucalyptus, cocoa and fruit. Funding will be pursued through applied-research grants (FAPESP PIPE, BNDES Garagem, EMBRAPA) and AgTech accelerators.
The end of the module is not the end of CropTrack — it is the start of the go-to-market: taking the solution back to the field, now as a product.
"We don't monitor harvests. We protect investments that take decades to grow."
- FAO. Global Agriculture towards 2050. Rome: Food and Agriculture Organization.
- FAO. The State of Food and Agriculture, 2025.
- UNITED NATIONS. World Population Prospects, 2024. New York: UN DESA.
- EMBRAPA CAFÉ / MAPA. Perdas por doenças e pragas na cafeicultura brasileira.
- CNC — CONSELHO NACIONAL DO CAFÉ. Propriedades cafeicultoras no Brasil.
- CONAB. Acompanhamento da Safra Brasileira de Café — Levantamento 2024/25.
- IBGE. Censo Agropecuário.
- GRAND VIEW RESEARCH. Coffee Market Size & Trends Report (CAGR 5.3%).
- AGROLINK / EMBRAPA. Pragas causam perdas de até R$ 55 bilhões à agricultura no Brasil.
- Y COMBINATOR. Requests for Startups — "AI for Low-Pesticide Agriculture", 2026.
- ULTRALYTICS. YOLOv8 Documentation. Available at: docs.ultralytics.com.
Supporting material produced by the authors, in modulo 15/CropTrack/:
docs/— full project documentation (Docusaurus): the sprint roadmap (Sprints 1–10), the Roadmap Validation (Expectation × Delivery) and the Conclusion.backend/·frontend/— the CropTrack platform source code. Run instructions inREADME_RUN.md.metrics/— model metric charts and reports (confusion matrix, per-class metrics, mAP).notebooks/— training and analysis notebooks.