HelpRush is launching in India first. Live in Srinagar today, expanding by waitlist demand. See where we're live →
Engineering
Most marketplaces claim verification. We publish the architecture. The six-layer system that decides which pro you ever see.
By Engineering, HelpRush · 23 April 2026 · 12 min read
Trust on a service marketplace is not a feature. It's the product. Get it wrong, the platform dies — gracefully if you're lucky, in a regulator press release if you're not.
Our architecture is filed with the Indian Patent Office (IN 202511047441) and runs in production today. Here's how the layers stack.
Aadhaar and PAN, captured inside the HelpRush Partner app. We originally built this on DigiLocker; we've since moved off. Today the pro's phone scans each document with Google ML Kit text recognition — the OCR runs on-device. The extracted data (Aadhaar number, PAN number, name) is matched against what the pro submitted on web. A mismatch rejects instantly; a match moves them forward. We get the extracted fields plus the encrypted scan for audit.
A smart selfie step uses Google ML Kit face detection with classification and tracking. It's a live camera read, not a photo upload — a printed headshot or a still on another phone doesn't pass the capture quality gate. The selfie becomes the avatar, and we lock it; any later change fires a Risk Engine signal.
This kills the simplest fraud — the impostor with a borrowed phone, the duplicate document, the stolen ID — at the gate.
Per-category quizzes, written by senior pros, expert-reviewed, regularly rotated. A plumber who passes the AC quiz can take AC jobs. A plumber who fails the AC quiz cannot.
Five stages per category: DECLARED (claimed but unverified — never matched), QUIZ_PASSED (passed the test), PROBATION (working through the first five supervised jobs), VERIFIED (cleared probation), EXPERT (sustained high performance). Multi-language — English, Hindi, Urdu — so pros quiz in the language they think in. Attempt limits vary by path: Professional 3 tries, Hybrid 5, Practical unlimited. Fail and there's a 24-hour cooldown before retake. Dispatch only considers QUIZ_PASSED and above.
First five jobs in a new category are watched. Customer thumbs-up/down feeds into SkillFeedback records tied to the job. Three thumbs-down in the window trigger a mentor review before the pro proceeds. Clear probation cleanly and the engine promotes you to VERIFIED automatically.
This is our slowest layer to game. A bad actor would need to clear native KYC, pass quizzes, and then maintain quality across five real jobs under direct supervision before they can do harm at scale. By then the human signals have done their work.
A live composite score (0–100), recomputed on every job, KYC event, and risk signal. Six inputs: verification level, KYC status, skill-verification breadth, risk-inverse, completed jobs, and rating saturation (with a 10-review floor so one angry customer can't tank a new pro). Visible on every public provider profile. Drives dispatch priority and tier eligibility.
Trust is the only layer with a feedback loop — it moves with what the pro just did. The other layers are gates; this one is gravity.
Ten distinct signal types feed a per-pro risk profile (OK → WATCH → FROZEN → BANNED): chat-PII leaks, call-PII leaks, customer reports, payment anomalies, geo revisits (patterns that look like fraud staging), device-fingerprint changes, OTP abuse, duplicate documents across accounts, identity linking across accounts (same phone + same device + different name = flagged), avatar unlock after verification. Cross a threshold and the account auto-freezes pending review. This is live in production today.
The identity-graph piece matters most: a sybil attack would need clean Aadhaars, clean PANs, clean phones, and clean devices — all distinct — for every account. The graph makes the attack expensive enough to bin.
For trades where a formal credential exists — a licence, a trade certificate, a degree, a professional membership — pros can add it. We admin-review each one and, once verified, surface it on the public profile alongside the SVE badges.
If trust is enforced by code, the code can be inspected. If it's enforced by call-centre humans, the customer takes our word for it. We chose the inspectable architecture.
Keep reading
Provider · 5 min read
Forty senior service pros in Srinagar just became HelpRush's first Assured cohort — re-verified, badged, and ready for the calls that come at the wrong hour. A note on what the badge means, why it took us months to issue forty of them, and who's coming next.
Field · 5 min read
Most platforms expect partners to figure the app out on their own. We don't. Notes from a Saturday training in western Srinagar — kahwa, kulcha, a whiteboard, and seven trades.