Biz Alırız

the problem

Turning real-world listings into one defensible number is a statistics problem before it's a product problem: naive averaging gets wrecked by a handful of scam-cheap or heavily-optioned outliers, and a "never overpay" guarantee has to hold even after that cleanup runs. Separately, uploaded vehicle photos routinely contain a visible license plate, which is personal data — routing every photo through a third-party vision API to redact it would defeat the purpose of protecting it.

how it is built

• Astro and React front end, FastAPI services, PostgreSQL (via SQLAlchemy 2 async, with Alembic migrations), delivered through Cloudflare.
• A valuation engine shipped as an importable Python package rather than a separate microservice: comparable-matching, statistical cleanup, condition-based deductions and liquidity-tiered margin all read from one YAML config that a test cross-checks against the written business rules.
• Adaptive reference price: the 25th percentile of clean comparables when there are at least 8 of them, otherwise the single cheapest clean comparable — both strategies are computed and logged on every run so the switchover point can be recalibrated from real outcomes later.
• A hard ceiling the offer can never cross (never price above the cheapest genuinely clean comparable), enforced on the raw comparable set independently of the statistical cleanup step.
• On-device license-plate detection: a YOLOv9 model (via the open-image-models ONNX runtime) runs on CPU inside the service, so uploaded photos never reach a third-party vision API.
• The plate is covered with a fitted, rotated brand panel instead of a blur: an HSV color mask (low saturation, high value) finds the plate's white against the vehicle body, with an axis-aligned box as the fallback when the fitted shape can't be trusted.
• Detection confidence threshold is set to 0.20, below the underlying library's own 0.25 default, on purpose — a wrongly-placed panel costs nothing, a missed plate is a data-protection failure.
• Cloudflare Turnstile gates the public lead form, and a background outbox worker retries failed webhook delivery with backoff instead of silently losing a lead.
• Python 3.12 workspace (uv), with ruff and pytest run as part of the workflow, not after the fact.

the hard parts

• Cleaning bad comparables (removing scam-cheap and heavily-damaged listings) has to run before a reference price is computed, but the safety ceiling — never price above the cheapest clean comparable — has to run against the raw, uncleaned cheapest comparable, or the two guardrails could contradict each other. The engine keeps a normalized set and a raw set side by side specifically so that can't happen.
• The detector's output is a loose axis-aligned box, but a plate photographed at an angle isn't axis-aligned. The service re-derives the plate's actual rotated quadrilateral from an HSV mask inside that box, discards the measurement if the recovered angle looks implausible (a reflection usually reads as a steep, wrong angle), and falls back to the safe axis-aligned box — coverage always wins over a tidy-looking panel.
• Matching comparable vehicles correctly turned out to hinge on which field on a listing "engine identity" gets read from: an earlier version pulled it from the wrong place and silently mismatched a meaningful share of listings against the wrong engine variant — a bug that would have quietly mispriced cars rather than crashing, caught only by cross-checking that field against the vehicle's own category path.