A Brazilian-Portuguese rebuild of Mike that can run entirely on your own hardware and wear your own brand.
This is @brauliogusmao's take on Mike, rebuilt for Brazil rather than merely translated. The whole interface has been adapted to Brazilian Portuguese with product decisions aimed squarely at the local legal market, and there's a Portuguese-language changelog documenting the work - a sign someone is treating this as a real product for real users, not a weekend experiment.
Two choices define the direction. First, it can point at locally-hosted AI models instead of only calling out to remote providers - meaning the model can live entirely on your own machine, no cloud round-trip. Second, a whitelabel layer lets each deployment set its own app name, logo, page title, and link-preview metadata, so the same product can ship under different brands. Together they read as a fork built for firms that want private, self-hosted legal AI they can put their own name on.
The rest of the work is the unglamorous stuff that separates a demo from something you'd deploy: surfacing real error messages when the AI fails mid-answer instead of a generic stream error, and fixing an upstream bug where live responses containing a table or formula could crash the chat before finishing. @brauliogusmao is worth a click if self-hosted, Portuguese-first legal AI is your thing.
i18ninfrastructurebranding
brauliogusmao ↗ analysis ↗ GitHub This fork can now point at locally-hosted AI models instead of only calling out to remote providers.
brauliogusmao added Ollama as a provider, routing `ollama/<model>` IDs through a new 309-line streaming adapter. The interesting part isn't the wiring -- it's that document-fetch tools are preloaded into context rather than called reactively, because small local models don't handle reactive tool use reliably.
A sweeping localization of the whole product into Brazilian Portuguese, bundled with a path to run the assistant on your own servers instead of a hosted AI provider.
A fourth AI engine option means the model can live entirely on local hardware - no cloud round-trip required.
brauliogusmao wired Ollama in as a fourth LLM backend alongside Claude, OpenAI, and Gemini. The key design decision: instead of letting local models call document-fetch tools reactively, the adapter preloads document content into the system prompt before the first LLM call.
When the AI fails mid-answer, users now learn why instead of staring at a generic "stream error."
When the LLM API returned an error during streaming, the chat UI would silently stop -- spinner gone, no message. brauliogusmao fixed the silent failure and then upgraded it to a typed error contract with six classification codes.
A disciplined, end-to-end Portuguese-Brazilian adaptation of the whole interface - with product decisions baked in for the local legal market.
brauliogusmao converted the entire frontend from hardcoded English strings to `next-intl` translation keys across 15 commits, targeting Brazilian legal market deployments. Every user-visible string from auth pages through tabular review now lives in `frontend/messages/pt-BR.json`.
A real upstream bug: responses that arrive live and contain a table or a formula could crash the chat window before they finished rendering.
Two commits fix a real `removeChild` exception in `AssistantMessage.tsx` that hits any React chat UI running `rehypeKatex` or `remarkGfm` while a stream is in flight. The fix took two passes to get right.
A whitelabel layer lets each deployment set its own app name, logo, page title, and link-preview metadata without touching the source.
brauliogusmao adds a thin branding layer that lets you swap the app name, logo, and HTML metadata via build-time env vars, with no code changes required. Six logo call sites are updated; spinners and status icons are left alone.
brauliogusmao · opened 2mo ago · merged 2mo ago by brauliogusmao brauliogusmao · opened 2mo ago · closed 2mo ago brauliogusmao · opened 2mo ago · closed 2mo ago