Muhammad Ramzan Aatish
Software dev: AI solutions for real business problems.
New to Contra
Muhammad Ramzan is ready for their next project!
I built a multi-tenant fintech operations platform. React on the frontend, Node API, Postgres with row-level tenant isolation underneath. It connects to Stripe, Plaid, QuickBooks and Xero through verified webhooks. Every event signature is checked against the raw request body, duplicates collapse into a durable idempotent inbox, and each event gets matched to the right tenant on the server before anything touches the ledger.
Money is stored in minor units, never as floats. The reconciliation engine catches amount mismatches, orphan payments and stale cursors. Anything it can't explain lands in an anomaly queue with a dead-letter lane. There's also an AI document pipeline, OCR plus extraction with prompt-injection canary tests, and an e-sign flow.
The test suite runs 325 tests green against a real Postgres. Dockerized, with database migrations and a disaster-recovery drill script.
0
6
I built a B2B lead generation app. You type in an industry and a location, and it finds real businesses through the Google Places API, then enriches them with company data and decision-maker emails via Apollo. Every email gets validated, and GPT-4o-mini writes a short personalized pitch plus a lead score for each lead. One click exports everything to CSV or Excel.
The three API keys are yours. You enter them once in the setup wizard and they stay encrypted on your own machine. Nothing is scraped, no proxies, no LinkedIn tricks, no SMTP probing. The app tells you exactly what each run costs in Google API fees before you spend it, filters out bad-fit leads before any paid enrichment runs, and picks up cleanly if the server restarts halfway.
It ships as a one-click Windows installer and a Docker image. React on the frontend, FastAPI on the backend, JWT auth, and 14 tests green.
0
12
I built an invoice extraction API that costs nothing to run. You upload a PDF or image, and it returns clean structured data: vendor, invoice number, dates, line items, subtotal, tax, total — every field with its own confidence score.
Under the hood it runs PaddleOCR on a regular CPU (Tesseract as backup), then a rule-based parser pulls out the fields. Before anything gets accepted, an arithmetic check makes sure subtotal plus tax actually equals the total — that catches the small digit errors OCR often makes. If something looks uncertain, it goes to a human review queue instead of failing quietly, and every correction is saved in an append-only log so you can always see what changed.
No paid vision APIs, no GPU needed. The whole stack is open-source, it ships in Docker, and it's covered by 28 unit and 4 smoke tests.
0
17
Built a Bring-Your-Own-Key AI agent platform in Python: clients connect their own provider API keys, so the platform never holds or pays for them.
It ships with an encrypted key vault (per-tenant, with rotation), a model router with per-call cost tracking and a global kill switch, a code-building agent that works inside isolated Docker sandboxes, and a WhatsApp control layer on Meta's official Cloud API with a full audit trail. There is also a research pipeline that verifies claims before they reach an article draft.
204 automated tests, all green. One command deploys it.
0
23