The Arithmetic Nobody Was Doing: Auditing E-Waste Recycling With Rules, Receipts and AWS
India's recycling certificates rest largely on documents recyclers upload about themselves. VeriCycle adds a witness at the first handover and runs sourced arithmetic on every claim, all serverless on AWS Mumbai.
The fraud was found by someone with a calculator
Look at how recent e-waste recycling scandals in India were uncovered. Nobody ran a sophisticated detection system. People did simple arithmetic after the fact.
- A news investigation found that 31 of 41 approved e-waste recyclers it examined looked non-existent or like ghost operations, licensed for up to 8.49 lakh tonnes a year (Newslaundry, July 2025). That was one outlet's survey, not a random sample, and we treat it that way.
- At a plant in Hapur credited with roughly 2,835 tonnes, inspectors reportedly found only dismantling activity (CPCB affidavit, 2 July 2025).
- A tax notice alleges that a photo stamped 29 minutes after sunset showed daylight (Reuters, 30 September 2026). These are allegations, nothing more.
Tonnes above plant capacity. Idle furnaces. A clock that disagrees with the sky. Every one of these is a check a patient person can do in an afternoon. The problem is that nobody does it on every claim, and nobody records the very first step of the chain at all.
The first handover is invisible
About 65 lakh e-rickshaw lead-acid batteries die each year in India. At roughly 20 kg of lead each, that is about 1,30,000 tonnes of lead a year. A large share moves through informal dealers, and the informal dealers often pay more than registered ones (around Rs 2,750 versus Rs 2,500 per battery in Delhi). So material flows toward cash, and a driver selling a dead battery to a dealer leaves no trace.
The gap, stated plainly: no tool checks a recycler's claim independently of the recycler's own uploads, and there is no record of the first handover.
What VeriCycle does
Vericycle is two halves that meet in the middle.
ChainTrack records the first handover with a 10 second WhatsApp voice note. A dealer says something like "chaar e-rickshaw battery, lagbhag 80 kilo, driver se li", the system transcribes it, extracts the numbers, reads them back for confirmation, and keeps a tamper-evident history of every later handover.
ClaimGuard runs nine rules on every recycler claim: capacity, mass balance, energy use, custody coverage, photo light versus time stamp, duplicate photos, document checks, weight drift and satellite site activity. It shows the working, the formula with real numbers filled in, and the cited source for every limit.
The one-line version: we do the arithmetic that used to happen after the scandal, on every claim, with receipts.
The rule that shaped everything: AI explains, rules decide
We deliberately did not build an AI fraud score. There is almost no labelled fraud data, and a black box cannot be defended to a recycler, a regulator or a court. A rule with a cited source can.
So generative AI is used for understanding (turning a voice note into structured fields) and explaining (answering "why was this claim flagged?"). It cannot raise or clear a flag. That decision belongs to fixed, versioned rules.
We wrote down eight principles and let them drive every screen:
- Consistency, not certification. A passing claim is "consistent with data", never "genuine". We issue no certificates.
- Missing data is not fraud. There are four outcomes: PASS, FLAG, INSUFFICIENT_EVIDENCE and INCONCLUSIVE.
- No source, no rule. Every limit cites a document. Invented demo numbers carry a visible DEMO VALUE badge.
- Three separate scores. Physical consistency, traceability coverage and evidence quality are never blended into one number.
- One-sided is never "witnessed". Each field records who said it and who confirmed it.
- Protect informal workers. No name, Aadhaar, GSTIN or address for sellers. Phone numbers are kept only as a keyed hash.
- AI explains, never decides.
- History is append-only. A correction is a new event pointing at the old one.
We also refused a few tempting things on purpose: issuing or trading certificates, calling anyone "fraudulent" (we say "inconsistent with data"), tracking single devices instead of batches, a refrigerant credit, and blockchain.
Why not blockchain?
We need tamper evidence, not agreement among parties who will never take part in a consensus network. KMS signatures, S3 Object Lock and a public daily digest give us that at a fraction of the cost and complexity.
The AWS shape of it
Everything runs in Mumbai (ap-south-1) and is serverless, so idle cost is close to zero.
Need | Service |
|---|---|
WhatsApp intake | AWS End User Messaging Social, SNS |
Voice to text | Amazon Transcribe |
Extraction and copilot | Amazon Bedrock (Nova Lite), Guardrails |
OCR on weigh slips and invoices | Amazon Textract |
Event store | DynamoDB on demand |
Tamper evidence | KMS asymmetric keys, S3 Object Lock |
Workflow | Step Functions, EventBridge |
Authorization | Amazon Verified Permissions (Cedar) |
Users | Cognito |
Front end | Amplify Hosting, CloudFront |
Infrastructure | AWS CDK v2 in TypeScript, cdk-nag |
The next three posts in this series go deep on the pieces: the WhatsApp custody ledger, the nine ClaimGuard rules, and how we kept generative AI on a short leash.
Where we honestly are
We would rather tell you now than have you find out later. The rule engine, review API, evidence processor, copilot, dashboard and the ChainTrack pipeline are built and covered by tests that run offline. The CDK infrastructure for both halves synthesizes and passes cdk-nag. As of this writing the stack has not been deployed to a live AWS account, and the WhatsApp number has not been linked. Several calibration steps (energy bands, the photo light classifier) need real plant data we do not yet have.
If a recycling-fraud tool cannot be honest about its own limits, it has no business judging anyone else's.
What we are hoping to learn
- Can an untrained dealer, comfortable in Hindi and not in English, log a handover in under a minute with no instructions?
- Will producers (brands facing disallowed tax credit for recycling that never happened) pay per witnessed kilogram?
- Which of our rules produce useful triage for human reviewers, and which need tuning?
Follow along. If you work on EPR compliance, reverse logistics or environmental audit, we would love to hear what is wrong with our assumptions.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article