top of page
Simple_Marble_Foundation_Logo__1_-removebg-preview.png

AI-Powered UXR Tools

Overview

​​​

A suite of 7 interactive tools built to accelerate every stage of the UX research lifecycle — from planning and data collection to analysis, reporting, and stakeholder communication. 

 

Tools were used and increased efficiency of a total of 14 quantitative research, 5 qualitative research, 2 data analyses, dealing with a total of 50,000+ samples, across all projects.

My role:

User Researcher

​

Skills

  • ResearchOps

  • Agentic Solutions (Claude, Qwen)

  • Workflow (n8n)

  • Vibecoding (Lovable, Cursor)

  • RAG & Knowledge Graph

  • ​Prompt Engineering

Time frame:

May 2025–September 2026

Text-Mining (Quant)

🐍 Python Text-Mining Pipeline (Quantitative)

Automating large-scale text analysis with NMF + LLM semantic clustering

Problem

  • Manual reading of open-ended survey responses across multiple languages is infeasible at scale

  • LLM-only analysis is expensive and inconsistent across large datasets

  • Traditional topic modeling lacks interpretive depth for nuanced insights

​​​​

  • Automated preprocessing pipeline with language-specific filters and minimum-length thresholds for 50,000+ responses across 16 languages.​

  • TF-IDF + NMF topic modeling paired with lexicon-first sentiment scoring and XLM-R transformer fallback for ambiguous texts.​

  • Qwen 3.5 performs LLM-driven semantic clustering on top — combining statistical rigor for pattern discovery with LLM reasoning for meaning-making

Solution

Results

82%

less time needed to complete analysis

47%

lower token cost vs LLM-only

16

languages supported

📚 Qualitative Coding Agent (Interpretive)

AI-assisted qualitative coding with full traceability and human-in-the-loop governance

Problem

  • Qualitative coding of interview transcripts takes days per study

  • AI-only coding lacks methodological rigor and auditability

  • Manual coding is slow and inconsistent across researchers

​​​​

  • Checkpoint-gated three-stage pipeline grounded in Saldaña, Strauss & Corbin, and Braun & Clarke — open coding, axial coding, selective coding — applied to 20+ transcripts (10,000+ words each).​

  • Claude Sonnet 5 as interpretive engine for long-context reasoning with mandatory human review between stages.​

  • Full quote-level traceability (transcript ID + timestamp) enforced at every stage; no code or theme is finalized without explicit researcher confirmation

Solution

Results

77%

less time needed to complete coding, from 9 hours to 2 hours

20+

transcripts processed

0

traceability failures

🎙️Interview Transcript Processing & Video Short Clips

End-to-end transcript processing with AI cleanup and instant video clip extraction

Problem

  • Interview transcript cleanup, review, and video clip extraction are fragmented manual workflows

  • Researchers switch between DOCX editors, SRT tools, and ffmpeg commands

  • Turnaround takes days per study across multiple studies

​​​​

  • Full-stack React SPA + PostgreSQL platform with format-specific adapters (7 timestamp regex patterns, speaker detection) processing transcripts in 16 languages

  • Multi-provider AI correction shown as word-level diffs with per-segment accept/reject​

  • ffmpeg-based caption burning with CJK-aware styling and instant clip extraction from coded timestamps via stream-copy with SRT verification — clips are self-contained, no separate subtitle files needed

Solution

Results

67%

less time needed for transcript cleanup

87.5%

less time needed for video clip extraction

52%

less time needed for end-to-end turnaround

📝 AI Research Plan Generator

Multi-model AI research planning with methodological guardrails and bilingual export

Problem

  • Research plan drafting is repetitive and template-driven

  • Different models excel at different tasks — using one model for everything produces generic output

  • Manual editing of AI-generated plans is tedious and time-consuming

​​​​

  • Multi-model pipeline with conditional routing serving 24+ projects in Chinese and English: Qwen 3.7 handles plan structure and translation; Claude Sonnet 5 handles question generation

  • Prompts enforce MECE options, bias avoidance (confirmation, framing, social desirability, Hawthorne), and past-experience grounding

  • Segmented markdown editor parses plans into 12 typed sections with per-section regeneration and a reusable question bank that prevents redundant generation

Solution

Results

52%

less time needed to create research plans

24+

projects served

12

plan sections auto-generated

❓ Research Question Generator & Optimizer Skills

Same-day UX research platform — from brief to visualized report in hours, not weeks

Problem

  • UXR request cycles take days to weeks — designers wait for researcher availability

  • Researchers are bottlenecked by operational work instead of high-impact studies

  • Self-service research without guardrails produces methodologically flawed instruments

​​​​

  • End-to-end platform: natural-language brief → structured instrument → UT deployment via MCP integration → API-based data fetching → AI analysis → visualized report, enabling same-day results for 40+ designers

  • 11 specialized AI flows across 5 frontier models (Claude Sonnet for question generation and semantic quality review, Qwen for topic modeling and 22-language localization, GPT-5.5 for intent extraction)

  • 27-check quality audit (12 structural + 15 semantic) catches violations missed during generation; all methodology constraints are non-negotiable prompt-level rules

Solution

Results

74%

less UXR operational time needed

40+

designers self-serving research

27

quality checks per study, as well as human-in-the-loop

🚀 Rapid Research Platform (AI-Powered ResearchOps)

Making self-service research scientifically defensible with executable methodology guardrails

Problem

  • Non-researchers (designers, PMs) need to write research questions but lack UXR training

  • Ad-hoc questions suffer from confirmation bias, hypothetical framing, and surface-level thinking

  • Flawed questions produce data that looks rigorous but isn't scientifically defensible

​​​​

  • Two complementary AI skills: Generator creates themed question sets from natural-language briefs with sub-theme clustering; Optimizer refines existing questions from text or uploaded files (DOCX/PDF/XLSX/CSV) with automatic type detection

  • Both produce structured bilingual markdown tables with purpose statements and standardized type labels

  • Quantitative prompts enforce MECE options, mandatory opt-outs, past-experience framing; qualitative prompts mandate scenario-based exploration and explicit bias avoidance — surface-level questions are systematically filtered in both modes

Solution

Results

2

complementary AI skills

24+

research projects supported

17

cognitive biases identifiable

©2026 by Arnaud Frattini.

bottom of page