MAD for Localization.
HACKATHON
Build the next useful localization prototype in 48 hours. A practical online hackathon where localization professionals define real-world problems, and builders turn them into working prototypes.
What happens
MAD Hackathon submissions
Explore the tools built by our participants during the hackathon. Test the live applications, see how they solve real tasks.
LocFit
by Egle Jasukaityte-Dziaute
LocFit is a rough-around-the-edges, first-pass triage tool that uses Claude to run HTML uploads against the MQM industry standard and flags the “must-fix” errors versus “minor polish”.
…
Show more
It’s not perfect – it only handles separate pages for now and sometimes trips over niche languages – but it does some of the heavy lifting and provides a smarter starting point than checking every string by hand. No matter the current shaky facade, the concept is sound with plenty of opportunities to build on.
Are We Fit for This Market (Website reviewer)
Model-Waldo
by Matteo Cehovin
Model-Waldo is a content-aware MT/LLM model selection engine for localization teams. It helps localization managers decide which current MT provider or LLM model is the best fit for a specific
…
Show more
translation request. The user uploads or pastes English source content, selects the target language, urgency, business priority, content type, and content requirements such as glossary support, placeholder preservation, or brand transcreation. Model-Waldo then analyzes the source content with an LLM, identifies localization risks, and ranks available MT providers and LLM models using explainable scoring rules.
The prototype recommends a specific model or provider, shows a strong alternative, explains why the recommendation fits the content, and indicates whether the localized output needs human review or automated QA before release. The goal is to move provider selection from a manual, inconsistent decision to a faster, more transparent, content-aware routing workflow.
Model-Waldo was built for the “Find the Right Model” challenge. It supports the challenge requirements by being flexible, automated, and explainable: users can adjust project constraints, the system performs the routing automatically, and the recommendation is supported by visible reasoning instead of a black-box answer.
Find the Right Model (Evaluation tool)
HeatWave
by CustoMWomen (Elena Murgolo)
An evaluation app for repeatable, scalable evaluation of new models.
The evaluation runs on challenge datasets, automatically created according to user specifications and integrated with
…
Show more
client data, using a variety of metrics, both rule and LLM based.
Outcomes can be reviewed by linguists and updated. Metric outcomes, model consistency and run times are displayed in the dashboard.
Models can be added at will and costs monitored in statistics.
Find the Right Model (Evaluation tool)
Optimal Model Selection Tool
by Lena Razumnik
Optimal Model Selection Tool helps teams choose the best LLM for multilingual translation by turning evaluation results into clear, explainable decisions. It integrates with LangSmith workflows by using LangSmith-shaped exports in the
…
Show more
current POC and is designed for live LangSmith API sync in production, so new models can be evaluated and reflected in rankings as soon as results are available. Users set business priorities across quality, latency, and cost, select language pairs, and instantly get transparent model rankings with score breakdowns and timeline context, making decisions faster, more objective, and easier to justify for both technical and business stakeholders.
Find the Right Model (Evaluation tool)
Website Localization Reviewer
by Andrea Danuzzo, Christèle Forget, Gary Hess, Sena Karaca, Estefania Tamayo Pineda
The Website Localization Reviewer analyzes translated web pages and scores them across four quality dimensions: Linguistic Quality, Terminology Consistency,
…
Show more
…Cultural Adaptation, and Completeness. Paste a URL and get an instant quality score — no knowledge of the target language required. For deeper analysis, provide the original source URL and the tool compares both pages directly, catching omissions and mistranslations invisible from the translation alone. Use the Multi-language tab to evaluate the same page in several languages simultaneously and compare results at a glance. Three analysis modes are available: AI Judgment, MQM Framework, and Comparison. Results can be exported as a text report or CSV.
Are We Fit for This Market (Website reviewer)
Context Minder
by Gabriela Janiszewska
Context Minder is a context-aware localization QA tool for the “Mind the Context” challenge. Instead of judging a translation in isolation, it reads the context around every string – string ID, category, length, tags, speaker, translator comments, plus your style
…
Show more
guide, glossary and character descriptions – and flags translations that are wrong for their context. It combines instant deterministic checks (length, tags, placeholders, plural/gender variants) with an LLM layer for register, terminology and meaning that auto-selects the right style-guide rules per string. Accepts any spreadsheet, runs cloud or local models, reports token usage, and exports a memoQ-ready report.
Mind the Context (Quality Pipeline)
MarketFit
by Vuk Lazovic
MarketFit answers one question for every market you sell in: is this website actually fit for this market, not just translated?
Paste any URL and MarketFit
…
Show more
automatically discovers every language version of the site (via hreflang), crawls each locale , rendering JavaScript when needed, and reviews them holistically. It scores each market on four weighted dimensions: linguistic quality, terminology consistency, cultural adaptation, and conversion readiness, combined into a single market-readiness score with a plain-English verdict (fit / needs work / not fit).
Are We Fit for This Market (Website reviewer)
Website LQA Auditor
by Christèle Forget
This Website LQA Auditor makes the quality audit of a live multilingual website significantly faster and more accessible than existing approaches, which typically require specialised
…
Show more
…tooling, manual linguistic review, or custom integrations. Given two URLs and a language pair, the auditor autonomously fetches, aligns, and scores website content against the TAUS DQF MQM framework and produces an actionable quality report in minutes — with no file preparation or toolchain setup. For higher-stakes content, the tool is designed to support rather than replace human judgement, with a built-in review and approval workflow. It has clear commercial applications: as a pre-sales diagnostic for LSPs demonstrating translation quality gaps to prospective clients, and as an on-demand quality gate for content teams before campaign launches or post-localisation updates.
Given two URLs, the auditor fetches, extracts, and aligns bilingual text segments, then applies the TAUS DQF MQM error taxonomy across 7 error categories — Accuracy, Fluency, Terminology, Style, Design, Locale Convention, and Verity — with severity-weighted penalty scoring (Critical = 10 pts, Major = 5 pts, Minor = 1 pt) calibrated to the TAUS B2B pass threshold of 50 penalty points per 1,000 words. The output is an interactive dashboard displaying a PASS/FAIL verdict, an A–F quality grade, a severity breakdown, an MQM category breakdown, and a priority matrix mapping flagged segments by visibility and conversion impact to help teams triage fixes efficiently.
The tool supports a full human-in-the-loop review workflow: reviewers can approve, modify, or reject AI-proposed corrections segment by segment directly in the interface. TMX synchronisation allows users to upload an existing translation memory file; the auditor identifies exact string matches between live website content and TMX entries, applies approved corrections, and exports a corrected, standards-compliant TMX 1.4 file ready for import into any CAT tool. For e-commerce sites, live currency conversion via the ECB Frankfurter API flags and corrects locale convention errors. The interface supports all 24 official EU languages, and the AI scoring engine accommodates multiple Gemini models, with a rule-based offline engine as a fallback for quota-constrained environments.
Are We Fit for This Market (Website reviewer)
Loc Sentinel
by Willian Magalhães
A ~99%-automated localization pipeline with exactly one human touchpoint — a single rubber-stamp in Slack. The model is kept on a leash: every translation is grounded in a per-market Source of Truth and verified by deterministic checks before a human ever sees it.
Show more
Localization Manager’s Dream (Capacity planner)
Relay
by Dinislam Zinnurov
Relay is a Slack-native intake and routing system for localization teams. Instead of forms or a separate ticketing tool, anyone — a requester, a product manager, a developer —
…
Show more
describes what they need in plain language, and Relay turns that message into a structured, tracked request and routes it to the right place.
At its core is a triage step that reads each message and extracts what matters: the type of work, target languages, deadline, priority, and whether the request can proceed or needs more detail. Because this classification is done by a language model rather than fixed rules, Relay handles request types it has never seen before — a routine file translation and a brand-new tooling integration are both understood with no per-type setup.
Flexible — No request types are hardcoded. The system reads intent and assigns a fitting category on its own, so the range of supported requests grows automatically as teams ask for new things. A request as unusual as setting up a glossary sync between two tools is classified and routed with zero code changes.
Human in the loop — Relay separates routine work it can queue directly from work that needs a person’s judgment. Complex or urgent requests are escalated to the loc team’s channel as an interactive card; a team member claims ownership and marks it done — keeping a human in control of anything non-trivial while the bot handles the busywork.
Ease of use — Everything happens in Slack, in plain language. Requesters get an immediate, formatted confirmation and can check their own request status anytime. The loc team works from a single live queue, and management sees all open work with one command. No new tool to learn, no context-switching.
Localization Manager’s Dream (Capacity planner)
RouteMT
by Andrés Romero Arcas
Pick the right MT provider — automatically, explainably
Show more
LoopLQA
by Kaja Braz
An end-to-end translation quality management system that automatically reviews multilingual websites for linguistic quality, identifies errors with AI, and learns from
…
Show more
human corrections through bidirectional Translation Memory (TMX) synchronization.
KEY FEATURES:
• Automated Quality Assessment – Ingests HTML pages, evaluates translations across 5 dimensions (Accuracy, Fluency, Cultural Adaptation, Locale Conventions, Terminology) using AI models, and generates detailed error reports with severity ratings and suggested fixes.
• Interactive Review Dashboard – Browser-like preview showing translations in context with color-coded error highlighting. Each error includes source/target comparison, explanation, severity level, and TMX suggestions.
• Bidirectional TMX Sync – Unique feature: not only reads from Translation Memory but writes approved corrections back to TMX, enabling the translation memory to learn and improve over time.
• Human-in-the-Loop Workflow – One-click approve/reject for AI suggestions, live preview updates, and batch rebuild that applies fixes to JSON files, regenerates HTML templates, and syncs corrections to TMX automatically.
• Easy Scalability – Add new locales in 2 steps: create a JSON translation file, run one command. No code changes needed. Currently supports 6 locales (Italian source + English, Japanese, Polish, German, Portuguese) with architecture ready for 50+ markets.
• Cost Transparency – Pre-evaluation cost estimates across multiple AI models help users make informed decisions before spending on API calls.
Built with Python, FastAPI, Pydantic, and BeautifulSoup4. Production-ready with comprehensive testing and automatic backup systems.
Are We Fit for This Market (Website reviewer)
Selection Engine
by OpenText L10N
Our Translation Provider Selection Engine automates provider/model selection in fast-changing environments where new models are introduced frequently.
…
Show more
We ingest newly available models daily from OpenRouter and continuously refresh benchmark data so recommendations stay current.
For each request (language pair, domain, and content type), the engine evaluates candidates across key dimensions such as quality, cost, and latency, applies configurable business rules (named profiles, custom weights, and constraints), and returns the single top recommendation.
This solution is:
Flexible: decision logic is policy-driven, so teams can adjust priorities (quality-first, cost-optimized, low-latency, etc.) without changing core architecture.
Automated: selection is computed end-to-end by the engine and delivered as one best result.
Explainable: every recommendation includes the score breakdown and decision metadata (active rule profile, weights, and benchmark version), making outcomes transparent and reproducible.
The result is a scalable, consistent, and auditable way to choose the best MT provider as model ecosystems evolve rapidly.
Find the Right Model (Evaluation tool)
ModelMatch AI
by OpenText L10N
ModelMatch AI is an intelligent translation provider decision engine that automatically benchmarks and evaluates multiple MT/LLM providers against real-world datasets.
…
Show more
By analyzing quality, cost, latency, language coverage, and business rules, it recommends the optimal translation model with clear, explainable reasoning. This enables localization teams to make faster, data-driven decisions while reducing evaluation effort, costs, and deployment risks.
Find the Right Model (Evaluation tool)
TermCaller
by Oskar Huledal
TermCaller extracts terms from callouts of technical illustrations to generate a visually-grounded glossary, importable as a TBX file.
Show more
By analyzing quality, cost, latency, language coverage, and business rules, it recommends the optimal translation model with clear, explainable reasoning. This enables localization teams to make faster, data-driven decisions while reducing evaluation effort, costs, and deployment risks.
Mind the Context (Quality Pipeline)
3 Days of Creation
- Challenges announced
- Watch the conference broadcast live
- Start building
- Build your prototype
- Use any AI-assisted development tool
- Connect with professionals on Slack
- Submission deadline
- One link or file per team
- Optional 2-min pitch
The Rules of the Localization Hackathon
- All code and prototypes must be built during the 48-hour window (June 24, 15:30 – June 26, 15:30 Paris time)
- Teams can pick 1 of the 4 announced challenges
- Open source libraries and AI coding tools are encouraged and expected
- Any AI-assisted development tool is allowed.
- Final submissions are published publicly for community review here
- Optional: 2-minute pitch slot on Day 3 for prize judging (all team members welcome)
Challenges of the Hackathon
“Mind the Context” (Quality Pipeline)
Context-aware translation remains a largely unsolved problem in our industry, and one that becomes increasingly important as more content is translated with AI. In this challenge, build an app that validates the translation in-context, on top of it being linguistically accurate. Pick your definition of context: developer notes, UI screenshots and HTML, dialogue characters and their personality descriptions, a video with frames taken every 2 seconds, images of products in an e-commerce catalog, PDF blueprints with formulas, or even an audio stream.
You can think of the quality pipeline as a quick and low-latency step that happens before translations go live, or as a post-publishing corrective process.
“Find the Right Model” (Evaluation tool)
In a world where new models become available every two weeks, some companies have opted to use all of them for their workflows. Different languages, tasks and content types require optimal models for characteristics like cost, latency, expected quality. Selecting new models has become so routine that it needs to be automated. Build a “Translation provider selection engine” that helps select the best MT provider. Requirements:
- Flexible: various business rules can be applied for decision-making
- Automated: humans should only be involved when required
- Explainable: decisions are motivated
Suggested by Nicolas Jadot, Trendyol
“Are We Fit for This Market” (Website reviewer)
Build an app that reviews and scores multilingual websites for linguistic quality, terminology consistency, cultural adaptation and other criteria. It would ingest HTML web pages, look at them holistically and generate a report or dashboard that helps content groups determine what to improve.
Advanced:
- Allow for the sync between the strings parsed from the website directly and strings residing in the translation memory, so that the app may find the exact position of the text within the TMX and fix it.
- Allow human review for proposed changes and an approval process
Suggested by Hyunjoo Han, Autodesk
“Localization Manager’s Dream” (Capacity planner)
A localization team is called upon to support multiple requests from various teams. Some are classic requirements such as file translation; others are more complex engineering jobs and involve checks across various tools, alignment between team members, or putting in place an entirely new workflow. The ask: build a planner/ticketing system tailored for localization and operating via a Slack bot that may support both the loc team and the stakeholders, including the requesters and the management.
Requirements:
- Flexible: can support an increasing number of use cases/request types automatically
- Human in the loop: knows when to loop in a team member
- Ease of use: stakeholders can interact with it directly and get useful answers
Suggested by Giulia Tarditi, Revolut