Eqaro
Four years running a fintech's platform: 5× growth, zero downtime.
The backend and web platform from the first commit, then the underwriting and claims agents on top
- Client
- Eqaro · Indian rent-guarantee fintech
- My role
- Tech head (consultant)
- When
- 2021–2026
- Status
- In production
Context
Eqaro guarantees rent for tenants, so landlords can skip the large deposit. Every application depends on outside KYC, credit bureau and payment APIs, and every claim is high-value money.
The problem
The platform had to carry fast user growth without going down, while every third-party failure (a slow KYC API, a vendor timeout) turned into a stuck application and a support ticket. Later, the team needed claims, underwriting and everyday reporting to stop depending on hours of manual work and on an engineer to run queries.
What I built
A modular Node.js backend: applications, KYC, underwriting, bonds, payments, claims and renewals. I wrote 1,612 of its 2,931 commits.
The Next.js customer and operations web app, with 929 of its 2,346 commits mine.
Retries, circuit breakers and idempotency on every KYC, bureau and vendor API, with monitoring and alerting.
Reads the verified file and recommends approve, reject or review, with a confidence score. Only borderline cases reach a person.
An orchestrator and 9 sub-agents move each claim through a 22-state process with 6 human approval gates and an append-only audit log.
Staff ask questions in plain language and get answers from live data through read-only queries, plus a 7 am founder brief.
The hard part
A claims agent reads emails, documents and bonds, then proposes a settlement. If the model ever did the arithmetic, one wrong number could reach a payment.
The model only reads and drafts. Every amount and date is computed by code, settlement needs two people to sign, nobody can approve their own work, and an approver can only lower an amount.
How it stays safe
A retried request has the same effect as one request, so nothing is charged or created twice.
Applications are decided automatically only at ≥0.85 confidence to approve or ≥0.90 to reject. Everything else goes to a person.
The WhatsApp assistant can't write to the database, and its results are capped at 500 records.
Aadhaar and PAN numbers never reach a model.
Result
5× user growth with zero downtime, 40% fewer KYC and vendor API failures, and 42% lower latency at peak. The underwriting agent is in production; the claims agent is being wired to production systems.
Stack
Start a similar project.
Find out in 48 hours exactly what it takes to get your app approved, live and earning.
Audit fee credited to your build · NDA on request
- A plain-English verdict: keep, fix or rebuild the risky parts
- Every issue ranked: what could lose you money or users first
- A backend and review check: data rules, exposed keys, payment webhooks, and Apple's and Google's top rejection reasons
- A fixed price to get you live
Credited to your buildFree · no call needed