01Waiting for the lab
A phone photo, coached until it is good enough.
- Level, marker, focus and light checked before anything is analysed
- A bad photo is refused with a reason, in Hindi or English
A grade at the gate, in seconds (target)
EnterpriseGrainGrade · wheat quality from a photo
One phone photo. A lab-style wheat quality report.
For farmers, commission agents, mandi labs, FPOs, flour mills and warehouses. Grain count, size in millimetres, defect percentages and the FCI FAQ grade — every number with an honest range, and a bad photo refused with coaching rather than analysed into a wrong answer.
- Under 30 s photo to report (target)
- 7 defect classes
- EN · हिंदी on the phone
WhySound familiar?
Three things we hear in every first conversation.
From mill buyers, mandi labs, FPOs and warehouse managers. Composite quotes from discovery calls — the problems are real; the names are left out.
-
“The truck waits while the sample goes to the lab.”
Manual grading takes a trained person, a sieve set and time — so trucks queue and decisions wait.
Mill procurement headHours per lot at the gate
-
“Two graders, two different answers.”
Counting broken and shrivelled grains by eye varies by person and by shift, and every disagreement becomes a price dispute.
Commission agentDisputes over every value cut
-
“We have no history of what each lot looked like.”
Results live on paper slips, so nobody can see whether a supplier or a stack is getting worse.
Warehouse managerNo trend, no evidence
HowWhat changes with GrainGrade
The same three problems, on the screens that solve them.
Product screens with illustrative data — GrainGrade is in development, with counting built and measured.
02Grading by eye
The same count and grade, whoever takes the photo.
- Every grain counted and measured; defects against FCI FAQ limits
- Every number shown with its range, never a bare decimal
One answer both sides can see
03No history
Every sample saved to its lot, with trends.
- Lots by source — mandi, FPO, stack, gate
- A shareable PDF or WhatsApp report for each sample
Quality you can trace over time
FieldGrainGrade in the field
The people, the phone, the place.
The workplace it runs in, the person who uses it on a phone, the team at the desk and the detail it reads.
Photographs are illustrative — real customer sites are not shown.
- 3.5%Counting errormeasured on our benchmark · target ≤ 3%
- ± rangeOn every numbernever a bare decimal
- 7Defect classesfrom broken to bore-holes
- FCI FAQRulepack built inplus mill and export specs
- 6Capture checksbefore any analysis
- 1 tapSharePDF or WhatsApp
GrainGrade is in development. Counting is built and measured; the other figures are design targets.
01The screens
From the camera to the lot report.
The guided camera, the result on the phone, the shareable report and lot history.
autowhat · GrainGrade / Guided capture
Illustrative data
The camera coaches the photo — level, marker, focus and light — before anything is analysed.
Product screens with illustrative data. The grain images inside them are drawn, not photographed.
02Real samples
Built on the photos people actually take.
Trays and plates, uneven light, chaff, mustard and chana mixed in, heaps and scattered grain — the pipeline is designed and tested against samples like these.
03The capture protocol
Accuracy is won before the photo is taken.
The camera coaches the shot instead of accepting anything. Six checks run before a single grain is counted.
- Top-down onlyThe phone must be within ±8° of level — a bubble overlay guides it. Oblique photos make near grains look 40% bigger.
- One layerGrains spread so none sit on another. A heap only shows its top layer — and the app says so.
- A scale in frameA printed 50 mm marker card or a known coin gives millimetres, sieve-equivalent sizes and weight-based percentages.
- Sharp and well litBlur and exposure are checked first; a bad photo is refused with a clear reason in Hindi or English.
- A plain backgroundA matte mat — printable from the app with a marker and a ruler — so wheat stands out from the tray.
- Never a guessIf the photo can’t give a trustworthy answer, the app asks for another one instead of producing a number.
04The vision pipeline
Eight steps, each one tested on its own.
Pure, repeatable functions: every analysis stores the exact settings and version used, so any result can be re-derived.
-
01
Straighten the photo
Rectify
The marker turns any angle into a true top-down view with even mm per pixel.
-
02
Find the grain
Segment
Wheat is separated from the tray by colour, not brightness — so reflections don’t leak in.
-
03
Split touching grains
Watershed
Each grain’s lit ridge becomes a seed; the dark crease between two grains keeps them apart.
-
04
Count three ways
Count + range
Three independent estimates; the published count comes with a range, flagged if they disagree.
-
05
Measure each grain
Features
Length, width, shape, colour and local darkness for every grain, in mm.
-
06
Classify defects
Rules, then models
Broken, shrivelled, damaged, immature, other grains, foreign matter, insect bore-holes.
-
07
Grade
Rulepack
Defect percentages checked against FCI FAQ limits — or a mill’s or exporter’s own spec.
-
08
Report & correct
PDF · WhatsApp
A shareable report; tap any grain to correct it, and corrections train the next model.
05Defects & grading
Seven defect classes. Graded against the rules you trade on.
Grading rules are data, not code — FCI FAQ first, then a mill’s intake spec or an exporter’s contract on the same schema.
- Broken / fragmentsSmall pieces with mostly free edges
- ShrivelledThin and light at normal length
- Damaged / darkWeathered or blackened over most of the grain
- Green / immatureA green tint across the grain
- Other foodgrainsMustard, paddy, chana — by size, shape and colour
- Foreign matterStones, straw and chaff
- Insect-damagedA dark bore-hole inside an otherwise normal grain
- Sound grainCounted and measured — the baseline every defect is compared to
| FCI FAQ wheat | Limit |
|---|---|
| Foreign matter | ≤ 0.75% |
| Other foodgrains | ≤ 2% |
| Damaged | ≤ 2% |
| Slightly damaged | ≤ 4% |
| Shrivelled & broken | ≤ 6% |
| Moisture (meter) | ≤ 14% |
What the camera cannot see: moisture, protein, gluten, aroma or insects hidden inside a grain. Moisture is entered from a meter reading — FCI grading needs it — and every report says so plainly.
06Where it stands
Counting is built. The rest is on a clear path.
Each phase ships only when it meets its measured acceptance bar on a hand-counted test set.
- Built · Scaffold & countingApp, API and the classical vision pipeline; counting measured at 3.5% error on our benchmark.
- Next · Guided capture & reportThe coaching camera, results screen, FCI rulepack and the PDF report.
- Then · Review & learningTap-to-correct, a golden test set gating every release, then trained defect models.
- Then · Weight basis & lotsWeight-based percentages, thousand-kernel weight, custom rulepacks, lot trends.
FAQQuestions we get about GrainGrade
Straight answers, before the call.
GrainGrade is a phone photo becomes a grain count, millimetre sizes, defect percentages and the FCI FAQ grade — with honest ranges — the same result whoever takes the photo, saved to the lot with trends by source. Short, factual answers for buyers, procurement and the assistants people ask first.
In short. GrainGrade is a phone photo becomes a grain count, millimetre sizes, defect percentages and the FCI FAQ grade — with honest ranges — the same result whoever takes the photo, saved to the lot with trends by source. It is built for mill buyers, mandi labs, FPOs and warehouse managers. It exists to remove the problems those teams describe in almost every first conversation: hours per lot at the gate while trucks queue; and value cuts disputed because counts vary by person. With GrainGrade, teams get capture coached until level, marker, focus and light are right, and every grain counted and measured; defects against FCI FAQ limits. It connects to Phone camera, Vision models, FCI FAQ limits and Lot ledger, is used by Buyers · labs · gate staff, and runs on mobile + web. Like every Autowhat AI product, GrainGrade goes live on one site, lane or process first, usually within weeks, on managed cloud, in your own VPC or on-premises, and shares one operational record with the other products, so the second deployment reuses the data and integrations of the first.
Who is GrainGrade for?
Mill buyers, mandi labs, FPOs and warehouse managers. A phone photo becomes a grain count, millimetre sizes, defect percentages and the FCI FAQ grade — with honest ranges — the same result whoever takes the photo, saved to the lot with trends by source.
What problems does GrainGrade remove?
hours per lot at the gate while trucks queue; value cuts disputed because counts vary by person; results on paper slips, with no trend.
What do we get with GrainGrade?
capture coached until level, marker, focus and light are right; every grain counted and measured; defects against FCI FAQ limits; every sample saved to its lot — mandi, FPO, stack, gate.
What does GrainGrade connect to, and who uses it?
It connects to Phone camera, Vision models, FCI FAQ limits, Lot ledger. Users: Buyers · labs · gate staff. Channel: Mobile + web.
How do we start with GrainGrade?
With a 30-minute working session on your own data — an engineer from the relevant sector, not a sales queue. The first wave is scoped in that call; most products go live on one site, lane or process within weeks and then scale.
Where does GrainGrade run, and who owns the data?
On managed cloud, in your own VPC (AWS, Azure or GCP) or on-premises, with your keys. SOC 2 Type II and ISO 27001 controls, India data residency when required, and local or open-weight models when policy demands it. The data stays yours.
07Pilot
Grade wheat at your gate or mandi? Help us test it on your samples.
Mills, warehouses, FPOs and mandi labs: send samples and your current lab results, and we’ll compare them side by side.