Visa support letter editor prototype
- Legal · Immigration
- 2026
- Full-stack engineer
- Cross-Border Document Service
E-2 employee support letter
Explore a fictional employer letter, review one contextual wording suggestion, and switch to a clean document view.
Sample only: every company, person, place, date, and circumstance below is fictional. Do not enter personal or sensitive information. Changes stay in this page session and are not uploaded or saved.
Sample letter ready. Your edits stay in this page session.
El problema operativo
I built a prototype for preparing visa support letters from applicant documents, with the full review flow in one workspace: upload files, inspect OCR results, generate a structured draft, edit it in a document editor, review contextual AI suggestions, and track revisions. The prototype was not adopted for daily operations because the client felt the generated text added too little beyond raw LLM output and subscription-based models were more cost-effective than API access.
Antes
- Source documents and letter drafting handled in separate steps
- Generated text reviewed outside a structured editing workflow
- Document changes difficult to compare and restore
Después
- 01Four-step flow from visa selection to draft generation
- 02OCR review and editable, sectioned support letter
- 03Inline AI suggestions with apply or dismiss controls
- 04Version preview, comparison, and restore
Lo que construí
- Built a structured document editor with direct editing, text highlights, and controls for applying or dismissing AI suggestions.
- Implemented version history with preview, comparison, and restore flows.
- Connected document upload and OCR review to the support-letter generation workflow.
- Integrated task-specific AI services: Upstage for OCR, Gemini for OCR summarization, Claude for letter generation, and OpenAI for review suggestions and interview preparation.
Antes / Durante / Después
Antes — hipótesis inicial
I set out to bring the document-heavy support-letter process into a single workspace, from source-file review through drafting and revisions.
Durante — restricciones y pivotes
I built a four-step intake flow and a document editor with contextual AI suggestions, text highlighting, and version history. Different AI services handled OCR, summarization, drafting, and review.
Después — resultado
The prototype was not adopted for daily operations. The client felt the generated text was too close to raw LLM output and that subscription-based models were more cost-effective than API access. The editor remained a meaningful implementation, but the end-to-end workflow did not prove enough differentiated value to justify rollout.
I would validate the generated content's incremental value and the cost model with reviewers before building the surrounding workflow. A capable editor matters when it supports an output users consider worth adopting.
Evidencia y alcance
El medio caótico — supuestos, revisiones, tradeoffs
- A complete workflow was not enough
The prototype brought upload, OCR review, drafting, and editing together, but the client still saw the generated text as little more than raw LLM output.
- API cost versus subscriptions
The client judged subscription-based models more cost-effective than paying for API-driven generation, so the prototype did not move into daily operations.
Tecnologías y técnicas aplicadas
Herramientas: React Router v7 (SSR) · Tailwind · Supabase · Upstage Document AI + OpenAI + Google Gemini + Anthropic Claude · PDF + DOCX parsers