At a glance
A leading West-African digital media organization, working on behalf of a major global health funder, needed to understand how women in Nigeria were actually talking about contraceptive products in their own words — at speed, across the platforms they used, and in their own languages. UNVEIL was retained as a subcontractor to design and build the AI analytics platform that produced those insights end to end.
The situation
Traditional research methods for understanding contraceptive attitudes — surveys, focus groups, formal academic studies — are rigorous but slow and expensive. They miss real-time shifts in public sentiment, regional and demographic nuance, and the unsolicited side of the conversation that surfaces in everyday social-media discourse.
The funder’s strategy increasingly emphasized user-centered, responsive research — listening to women in their own voices, in close to real time, to inform product investment, communication, and market strategy decisions. Social media offered the velocity, scale, and authenticity to complement traditional research, but harvesting useful signal from it required a serious AI pipeline.
The challenge
Several constraints made this a hard problem:
- Multi-platform, multi-format. The conversation lived across Facebook, X (formerly Twitter), Instagram, YouTube, TikTok, and a major Nigeria-specific community forum — as text, images, and video, with culturally important content on the local-forum platform that global research tools largely ignore.
- Multilingual. Posts mixed English, Nigerian Pidgin, and code-switched local dialects. A simple English-only pipeline would miss most of the meaningful signal.
- Privacy and ethics. Even working with publicly visible content, the analysis had to anonymize and deidentify aggressively to respect the people whose conversations were being studied.
- Three analytical lenses, one pipeline. Sentiment tells you the feeling. Content analysis tells you what topics are present. Thematic analysis tells you the meaning. Each is a distinct discipline; the funder needed all three, integrated.
- Production-grade, not one-off. Beyond an initial report, the funder wanted an automated pipeline and a live dashboard — something that could keep producing monthly and quarterly insight without re-doing the engineering each cycle.
Our approach
We delivered the system as a real software product, not a consulting deck.
- Multi-platform ingestion. Built data-collection adapters across the relevant social platforms — including a Nigeria-specific community forum — using commercial APIs and platform-appropriate scraping.
- Format-aware preprocessing. Text was normalized; videos were transcribed via speech-to-text AI; image-based posts had their text extracted via OCR; non-English content (including Nigerian Pidgin) was language-detected and translated to a common analysis language.
- Privacy by design. All content was anonymized at ingest. Personal identifiers, hashtags, URLs, and direct quote attributions were stripped via a combination of regex rules and AI assistance before anything entered the analysis layer. Access was scoped under least-privilege principles.
- Three-method analytical layer. Built sentiment classification using transformer and LLM models; content analysis with an LLM-driven coding pipeline against a versioned codebook (with human-in-the-loop review and intercoder-reliability statistics like Cohen’s Kappa); and thematic analysis with topic modeling and word-cloud summaries.
- Cloud-native deployment. Deployed on AWS — object storage for raw and processed corpora, serverless compute for NLP workloads, a managed PostgreSQL database for processed results, and an interactive analytics dashboard. Authentication and access control were enforced through cloud-native identity services.
- Live dashboard. Stakeholders could filter and explore findings by platform, product, theme, code, and time period — and pull monthly or quarterly summary reports without us in the loop.
The outcome
- A working, deployed analytics platform that turns ongoing multi-platform, multilingual social-media discourse into structured insight on the conversation around women’s health products.
- An initial comprehensive report, a versioned codebook manual, a curated domain-specific search vocabulary, documented prompts and validation procedures, and an automated pipeline that re-runs without manual intervention.
- A live, stakeholder-facing interactive dashboard with role-based access — replacing slide-deck deliverables with a continuously updated decision-support tool.
- The system has been running autonomously since deployment, generating recurring monthly and quarterly insights with minimal ongoing engineering.
What this means for you
If you have a customer base, a public, or an audience whose authentic, unsolicited voices live in social channels — and you want to understand what they actually think, in their actual languages, faster than traditional research can move — we can:
- Build a production-grade social-listening AI platform that combines sentiment, content, and thematic analysis on top of the platforms where your audience lives.
- Handle multilingual, multi-format content including text, images, and video.
- Engineer privacy and ethics in from the start — anonymization, deidentification, access controls, and documented data-retention policies.
- Deliver an interactive live dashboard, not a one-off report — so insight keeps compounding without re-paying for the engineering.
- Do this on your cloud account, with your security posture, so you own the data and the system.
Want to explore a similar problem? Contact us.