Research made usable
Posts and attachments are processed according to their structure, with links back to the original sources.
← Back to Projects
Selected case study · Cultural strategy agent
Problem
Solution
01
02
03
04
Value
Ask Collage
Ask Collage answers questions about cultural strategy, marketing and audience behavior using brand research, ad tests and consumer research. It combines research content, quantitative analysis and conversation context, with citations and related content clients can explore.
FocusResearch synthesis and quantitative analysis for strategic decisions
ResultMain product-interest driver for clients; 120% of the Q1 new-business goal
Years of research, spread across reports and data.
Research lived across posts, PDFs, presentations, webinar transcripts and survey data. The goal was to let clients ask a question and receive a useful answer without first knowing which report to read, which data to use or how to calculate the relevant comparisons.
Combining research retrieval and data analysis.
I designed and built the retrieval, analysis and recommendation workflows behind Ask Collage. The workflow depends on what the question requires: research content, analysis of survey data or a combination of both.
The RAG pipeline processes posts and attachments according to their structure. For webinars, I developed an algorithm that uses semantic similarity to identify changes in topic, so related discussion stays together when retrieved. Hybrid retrieval combines semantic and lexical search, recency weighting and reranking to select relevant research.
When a question requires quantitative analysis, a planner identifies the data source and metrics. Routing logic then determines which tools the analyst can use to retrieve data or calculate benchmarks, rankings and audience comparisons. The final response combines the findings with relevant research and conversation context.
Webinars organized by topic
Transcripts are segmented at topic boundaries so each retrieved passage preserves the context of the discussion.
Relevance with context
Retrieval selects relevant research, while analytical tools calculate the metrics and audience comparisons needed for the question.
A path beyond the answer
Citations and related content connect the response to reports, survey questions, audience segments and ad creative.
Metric calculations and source references
Metric calculations use established definitions and the appropriate respondent base. Minimum sample requirements determine which results can be included, and comparisons use indices or percentage-point differences according to the metric. These rules keep the calculations consistent across questions.
Related content search uses relevance filters, deduplication, LLM judging and canonical term matching to find the research behind an answer. I built citation processing and deep links so clients can open the relevant reports, survey results and ad creative.
I connected these workflows through a reusable LLM framework and an authenticated FastAPI service, with conversation memory and suggested follow-up questions.
Direct access to research and audience comparisons.
Clients can ask a strategic question, receive a synthesis of the relevant research and data, and follow the references to investigate further. Ask Collage became the main product-interest driver for clients and helped the business achieve 120% of its Q1 new-business goal.
Experience
Continue exploring