HTML-aware translation pipeline that never silently drops content
Problem
Site owners with 1000+ templated HTML pages fed to LLMs for translation get back pages that look translated but are silently missing whole sections — paragraphs summarized away, content dropped with no warning, and naive chunking breaks tags mid-element. Longer pages make it worse.
Opportunity
A DOM-aware translation pipeline that structurally splits pages at safe boundaries, translates while preserving markup integrity, and verifies nothing was lost (diff-based completeness checks) before publishing.
Market analysis
The failure mode is real and well documented: LLMs summarizing instead of translating and chunking that breaks markup. The white space is not translation quality but verifiable completeness, yet the broader market already has strong formatting-preserving players.
Market · Localization teams, agencies and site owners localizing large templated sites into multiple languages; steady B2B demand driven by international SEO.
Pricing · Comparable document-translation tooling runs from DeepL API plans (~$25/mo entry) to per-seat TMS subscriptions; per-page or per-character SaaS pricing is the norm and willingness to pay is real in B2B.
Pros
- + Verifiable, binary promise ('nothing lost') is a sharper selling point than translation quality.
- + DOM-level splitting at safe boundaries is a tractable engineering problem for a solo builder.
- + Fits CI pipelines: run as a build step that blocks publishing on failed completeness checks.
Cons
- − Crowded adjacency: DeepL document translation and enterprise TMS already preserve formatting.
- − Quality verification is semi-open: diff checks catch dropped content but not subtle mistranslations.
- − Buyers in localization are conservative and slow to switch pipelines.
Source
Hacker News (Ask HN)
The differentiator worth building is not the splitter but the verification contract: checksum every text node in the source DOM against the output before publish, and refuse to ship on mismatch. “Silent content loss” is the exact fear that keeps localization managers up at night, and no general-purpose translator sells that guarantee. The risk is that model vendors keep improving native markup handling until the pipeline becomes a thin check layer, so the completeness verifier should be the product, not the translator.