Research Prototype Human-AI Collaboration Multi-Agent System

Agent-Assisted User Research & Decision Support OS

A web-based multi-agent prototype exploring how LLM agents can assist B2B user research workflows, behavioral diagnosis, and evidence-grounded decision-making — without displacing analyst judgment.

Motivation

User research insights are often fragmented across qualitative interviews, behavioral data, and stakeholder intuitions. Decisions get made without traceable evidence. Can agents help structure this?

Research Questions
  • → How can agents support sensemaking without overriding user judgment?
  • → What workflows preserve decision ownership and traceability?
  • → How should conflicting signals across data types be surfaced?
Design Approach

A routing orchestrator delegates to scoped specialist agents. Each returns reasoned output, not just results. High-risk actions require human confirmation. All decisions are logged and traceable.

Status: Research Prototype (v1) · Not production software GitHub
Voice of Customer · User Research
Insight

Sentiment Monitoring & Competitive Analysis

Brand monitoring Competitor benchmarking
Own-brand sentiment tracking
Competitor model comparison
Industry user needs mapping

Tracks domestic and international sentiment via NLP clustering of key topics across smart terminals, EVs, and consumer electronics. Supports product planning and GTM decisions with evidence-cited outputs.

Research interest How should agents handle contradictory signals across platforms? When should they surface uncertainty rather than synthesize?
View Analysis Dashboard
Insight

Qual + Quant Research Agent

Pre-sale diagnosis Research plan Evidence chain
We need to improve B2B conversion — where should user research start?
Pre-sale diagnosis → Qual+Quant plan → fix findings to Decision Memory (Trace Graph).
Deliverables: interview guide · survey draft · metric definitions · evidence chain

Routes B2B diagnostic inputs into structured research plans. Outputs are citation-grounded and stored in a Decision Memory trace graph that constrains future decisions — preventing "evidence laundering."

Research interest Can a decision-memory architecture reduce the risk of agents generating plausible-but-unsupported recommendations?
Open User Research Agent
Service System

AI Customer Service: Consumer & B2B Optimization

Guided dialogue Compliance
I want to buy car insurance — how do I choose?
2 quick questions: new or used car? Any loan or commercial use?
Multi-turn clarification → plan recommendation → compliance note

Consumer: RAG knowledge base from major insurers; multi-turn dialogue narrows to explainable plan recommendations with compliance flags.
B2B: NLP + RAG complaint classification, root-cause attribution, and risk alerting — dashboarded for quality review.

Consumer ChatBot (Insurance) B2B Complaint Copilot
Behavioral Analytics & Data Analysis
A/B Testing

Advertiser Chain White-box Experiment Console

Running Guardrails Diagnosis
Total
24
Watching
5
Alerts
1
Unified view: results · guardrail risks · bottleneck diagnosis

Unified experiment workbench for ADX/SSP/DSP: metric lift, guardrail variance, funnel bottlenecks, and automated post-mortem notes — all archived in one place.

Research interest How do analysts decide to override or trust agent-generated experiment conclusions? What triggers confirmation-seeking behavior?
Experiment Console Advertiser White-box Dashboard
Lifecycle

Lifecycle Diagnosis & Automated Retention Agent

Diagnoses funnel drop-off points via user journey and lifecycle segmentation. Layered user models drive automated outreach (push / EDM / SMS) with vouchers, points, and time-limited incentives calibrated to segment and channel.

Research interest When agent-defined segments are used to target users, does analyst interpretability of segment logic affect decision confidence?
Lifecycle Retention Copilot
Attribution

Multi-Touch Channel Attribution Dashboard

Time-decay multi-touch attribution model validates incremental conversion contribution of AI-assisted shopping via switchback window experiments and channel-level inquiry/conversion tracking.

Research interest How should agents communicate attribution uncertainty to non-technical decision-makers without inducing over-confidence?
Channel Attribution View
AI Applications · Multimodal Generation & Evaluation
AIGC · Multimodal

Dance Flow: Human-in-the-Loop AI Interaction

Reference video Pose estimation Gemini
I want to learn this dance but can't keep up with the tempo…
Detected offset: right arm angle insufficient → suggest 0.75× speed + 3 repetitions.

Uploads a reference video to extract skeleton keyframes; webcam aligns and scores user movement in real time. Gemini generates corrective feedback and practice recommendations. Pipeline: frame extraction (0.1s) → pose alignment → similarity scoring → adaptive slow-down → review suggestions.

Research interest How does real-time AI feedback shape user agency in embodied skill acquisition? Does it augment or replace self-monitoring?
Google Hackathon Devpost
AIGC · Reflective AI

What-if: Reflective AI for Life Decision-Making

Key events Causal mapping
Stuck on career vs. city vs. relationship — feels like every choice forecloses another…
Let's map the key decision nodes first and run alternative what-if paths — who else's perspective matters here?

Transforms user narratives into structured decision scenarios with roles, causal chains, assumptions, and alternative what-if paths. Role-based agents (counterpart, observer, future self) support perspective-taking and reduce decision foreclosure. Memorial Award, Memory Genesis Competition 2026.

Research interest Do role-based agents increase perceived decision ownership? How does perspective diversity affect reflection depth?
Open What-if Prototype
Creative

Creative Insight & Content Generation Pipeline

Sentiment Insight
Hot keywords · complaints · reputation shifts
User Research
Survey · interview · segmented persona → scene motivation
Creative Insight Library
Tag decomposition · eval set generation · prior + posterior calibration
Script Generation
One-click script → multi-platform distribution

Distills scene-level user needs and psychological motivations from sentiment + research. Generates labeled evaluation sets for creative variants; prior scoring + small-traffic posterior calibration selects top-performing material templates.

Research interest Does agent-generated creative framing narrow the search space for human ideation, or does it anchor creativity too early?
MECE Combination Dashboard

Limitations & Scope

  • → Research prototype, not production software — routing is prompt-based, not learned
  • → No formal user study conducted; evaluation dimensions are design-intent only
  • → All demo scenarios are synthetic or researcher-constructed
  • → Optimized for exploratory demonstration, not latency or reliability

Related Work

  • → Sensemaking frameworks (Weick 1995; Russell et al. 1993)
  • → Human-in-the-loop AI design (Amershi et al. 2019)
  • → Decision support & traceability (Klein et al. 2007)
  • → LLM agent architectures (Park et al. 2023 — Generative Agents)