Saraschandra KaranamHuman - AI Interactions Researcher · Enterprise AI
Back to Research
RESEARCH PROGRAM • 2025-26

AI-Augmented Research

Experiments with GenAI applications to UX research processes

Search: Ask questions across feedback sourcesAsk me anything about customer feedback...SearchWhat do customers think about new navigation?Show me dashboard navigation issues in the last weekREACT Single Page AppSearch InterfaceSearch ExecutorAI Output SynthesizerMCP ClientOllama ClientHTTP / SSE / POSTOllama: Mistral 7BNode.js serverJSStdinStdoutFilesystem MCP serverNPSCSATSupport Tickets
Representative Visualization

One Knowledge Hub

System architecture diagram integrating multiple customer knowledge repositories into one interface using multiple MCPs.

4Experiments
500+JTBD comparisons
300+Highlights analyzed
3Customer feedback channels
02 · WHY THIS RESEARCH MATTERED

Scaling UX Processes with LLM-Powered Workflows

As a research study's scope, number of interviews/transcripts to encode and the number of stakeholders collaborating increased, I began building LLM-powered workflows to scale processes without sacrificing rigor and learn applications of GenAI to UX research in the process. This also enables UX research teams to align with the enhanced speed of execution by PM / Engineering teams with use of AI.

03 · REPRESENTATIVE CASE STUDY

Ask the Report

Background

Often, cross-functional stakeholders have many unanswered questions during the readout or follow-up clarifications after the readout. UX researcher is not always available, may be in a different timezone or busy with another project.

Research Questions

  1. Would it be possible to create a simple chat-based application that answers any questions on the research report?

Research Approach

To ensure qualitative rigor, I designed a four-stage pipeline:

  • CalibrateThe researcher manually codes the first 3 transcripts to define the tagging taxonomy.
  • Quality ControlThese human examples are used as few-shot training for the LLM. The script executes, and the researcher audits the tagging boundaries, refining prompts.
  • SynthesisThe script calculates tag counts and validates them against raw evidence before committing.
  • Ask the ReportVerified highlights and scrubbed transcripts are loaded into the vector index as grounded system context for stakeholder Q&A.
Few-shot promptingPII scrubbingScope guardlocal LLM for privacy
Ask the Report [ALPHA/EXPERIMENTAL]Q: What moment had the most missing features?Analyzing 8 transcript files...Answer: Participants repeatedly flagged the lacking filter options...Slides (HTML/JS)Ask the Report PanelPOSTSSE (Server Sent Events)JSNode.js ServerStatic Serving/api/askCorpus LoaderIn-house LLMClaudeOllama
Ask the Report: GenAI enabled. Enables cross functional stakeholders to ask any questions on a research report
04 · PROGRAM SYNTHESIS

What We Learned

Learnings from this program

Calibration beats instruction

Simply writing rules in a prompt is not enough; you must feed the model concrete, human-coded examples.

The bottleneck shifts, it doesn't disappear

While the AI reduces tagging time, the human researcher must spend more time designing validation schemas and auditing outputs.

Design the system first

Building a local model architecture with structured outputs is essential before running code

05 · RESEARCH ASSETS

Research Artifacts

Assets and frameworks generated to guide other UX researchers on the team:

Methodology

Re-usable methods for each experiment

Prompt Templates

Prompts used

Guidelines & Learnings

Best practices document

06 · EXPERIMENTAL TRIALS

Experimental Trials

A portfolio of experiments exploring generative AI capabilities across qualitative tasks.

Study 01

Experiment 01: Secondary Research

Outcome: Summarize past research reports in a pre-defined structure
Study 02

Experiment 02: Ask the Report

Outcome: AI-assisted coding + queryable deliverable — stakeholders interrogate findings in natural language during and post-presentation
Study 03

Experiment 03: Figma -> Study Guide

Outcome: Auto-generate research study guides directly from Figma design specs via Figma MCP
Study 04

Experiment 04: Calibrate-then-Replicate

Outcome: Given a large presentation (100+ slides), generate publication-ready structured insights using calibrate-then-replicate strategy
Study 05

Experiment 05: Compare personas

Outcome: Compare JTBDs against multiple personas with multi-method bias detection and correction
Study 06

Experiment 06: One Knowledge Hub

Outcome: Integrate multiple customer knowledge repositories into one interface using multiple MCPs
07 · PROGRAM OUTCOMES

Research Impact

The structural, organizational, and methodological contributions generated by this program.

Organizational Alignment

  • Best practices and learnings

Methodological Value

  • Prompt templates