Ferhat RudvanoğullarıData Engineer

Work

16 projects, 15 of them shipped since March 2026: pipelines, data quality and AI systems at SupplyBridge, plus a few of my own. Every number comes from a run log, a report or the commit history.

Selected

Supplier Page Classification

Located half a million suppliers' websites in the Common Crawl archive and classified their pages into seven categories: 485,562 domains in under 25 minutes.

domains in 24m 43s
485,562
domains in 24m 43s
pages in 7 categories
1.08M
pages in 7 categories

LLM Supplier Enrichment

Turned raw supplier pages into structured profiles with an open-source LLM running on our own GPU, and more than doubled its throughput.

domains/s, 1 process
2 → 4.8
domains/s, 1 process
domains/s, 3 processes
~8
domains/s, 3 processes

Address Normalization

Replaced a paid AI call per address with a self-hosted ML address parser (libpostal) and took throughput from 100 to over 15,000 addresses a minute.

addresses per minute
100 → 15,000+
addresses per minute
est., 1.07M addresses
~35 h → ~1.5 h
est., 1.07M addresses

Procurement AI Service

Turned a rule-driven procurement agent into a fixed pipeline behind a private API on Cloud Run, with a predictable output, capped model calls and evidence for every claim.

model calls/company
≤ 2
model calls/company
token cap, in / out
32K / 8K
token cap, in / out

Data pipelines

Built an 11-step pipeline for Vidyon that finds German solar installers, pulls their contact details with an LLM and scores every website on PageSpeed.

1,979 target companies, 98.3% in the right segment

Found and geocoded where suppliers actually manufacture, adding 59,791 verified locations across 56 countries.

59,791 new locations for 9,964 suppliers

Data quality & master data

Matched duplicate company records across 1.2 million rows with tiered rules, from exact website matches down to semantic similarity, with a human in the loop for anything uncertain.

1.2M company records

Checked the websites of 43,969 suppliers and repaired the wrong ones with web search and LLMs, cutting the manual review queue from 5,213 to 14.

5,213 → 14 records needing manual review

Safe Bulk Import

Jul–Aug 2026

Loaded 2.6 million product records into a live database under three rules: only fill empty fields, never delete, and never write anything ambiguous.

2,627,085 product records and 56,744 customer links

Cross-checked 23,443 facts about 140 major automotive suppliers against live web sources, with zero failed runs, for $0.37 a company.

140/140 companies audited, zero failures

Application development

Added Stripe subscriptions, expiring credits and consent-first analytics to a B2B sourcing platform so it could launch as a paid product.

Free → paid trial turned into a paid product

API Migration

May–Jun 2026

Moved an admin API from tRPC to layered REST and checked every endpoint against the old one automatically; the slowest call went from 14 s to 0.9 s.

43 admin endpoints migrated (out of 551 procedures inventoried)

VizCad

2025–now

My own product: a browser-based 3D viewer for storing and sharing large CAD files, running serverless on Cloudflare.

0 installs needed to open a CAD model

Deployment & DevOps

Set up automated deploys, a reverse proxy and one-command rollbacks, then moved every app onto a single server.

1.16 GB → 64 KB Docker build context

Put logs, metrics and request traces in one dashboard, then hit the system with spike tests to find out where it breaks.

0 → 300 concurrent users in 5 seconds (spike test)

Open source

Tried five different approaches to finding duplicate companies in 477,600 records, and wrote down why each one was dropped.

477,600 raw records, whole pipeline in ~8 min

Smaller projects

  • cad-pipeline Architecture for the VizCad backend that turns CAD files into GLB for the browser.
  • smelt GitHub Actions that set up the Daml SDK and build smart-contract projects in CI.
  • Hunyuan3D on RunPod Tencent's open-source image-to-3D model packaged to run on RunPod's serverless GPUs.