ComplaintsCompass

Inspiration

For consumers, knowledge is power, but unfortunately, it’s often out of reach. Many companies aren’t transparent, and crucial information about how to resolve an issue can be hidden or difficult to access. We noticed that complaints often don’t reach the right department, leading to internal misrouting, unnecessary delays, and dissatisfied customers. This inefficiency is especially harmful in industries like finance, where unresolved issues can have serious legal or financial consequences. We also observed how delays in customer support and vague responses add to users' frustration. We built ComplaintsCompass (powered by our core engine ClariRisk) to solve these pain points: intelligently routing complaints, assessing urgency, and guiding users toward faster, more effective solutions.

What it does

ComplaintsCompass is an AI-powered assistant that understands complaints in natural language and offers actionable next steps. When a user submits a complaint, the system identifies the key product, issue, and company involved. Using a vector search engine, it retrieves similar complaints from a large real-world dataset and analyzes them for red-flag keywords, legal signals, product severity, and repeat behavior. A risk score is calculated to prioritize high-stakes or time-sensitive cases. Based on these insights, the assistant suggests tailored actions—whether that means contacting a legal authority, sending a follow-up to customer support, or filing an official complaint—with ready-to-use templates and contact guidance.

Who is it for

• Consumers who are frustrated by opaque complaint processes or delays and want to resolve issues without doing exhaustive research or legal digging.
• Companies looking to intelligently triage incoming complaints, escalate urgent ones, and reduce customer support load.
• Governments and regulators aiming to monitor trends in public grievances, identify systemic problems early, and ensure fair outcomes for citizens.

ComplaintsCompass is designed as a SaaS solution, making it easy to adopt across multiple organizations and domains. Companies or institutions can deploy the assistant using their own complaint datasets or tune it for domain-specific applications with minimal setup. By simply changing the underlying data and configuration, users can tailor ComplaintsCompass to fit different languages, jurisdictions, or industries.

How we built it

We started with the open-source Hugging Face consumer-finance-complaints dataset, which contains over 500,000 real-world complaints submitted to the U.S. Consumer Financial Protection Bureau (CFPB). We cleaned and enriched this dataset with structured metadata (product, issue, company, date) to make it more useful for intelligent retrieval and analysis.

We used TiDB Cloud both as our primary database and to store dense vector embeddings for semantic similarity search. This allowed us to unify data storage and retrieval logic within a single, scalable SQL-based system. While we initially experimented with Qdrant, we streamlined the architecture by consolidating data and vector operations into TiDB to reduce complexity and improve maintainability.

Our backend is built in Python, orchestrated with LangGraph to manage a multi-agent workflow. These agents perform steps such as classification, search, risk analysis, and response generation. We use OpenAI APIs for language understanding and drafting resolution templates. The frontend is built in Next.js, providing a clean user experience for complaint submission, insights viewing, and letter drafting.

Challenges we ran into

• Designing the risk engine: We had to balance complexity and clarity. Combining red-flag detection, legal urgency, repeat offense signals, and product severity into a single score required iteration and explainability.
• Improving semantic search: Early versions returned many irrelevant matches. We added filters, metadata constraints, and deduplication to improve precision.
• Cross-system integration: Getting TiDB, Qdrant, LangGraph, and OpenAI to communicate reliably meant resolving data format mismatches, auth flows, and timing issues across the pipeline.

Accomplishments we're proud of

• Built a full-stack, production-grade assistant that goes beyond a simple chatbot.
• Created a meaningful risk scoring model that surfaces dangerous or ignored complaints with clarity.
• Used real-world data to create explainable, user-centric results.
• Designed a system that can scale from consumer use to enterprise or government adoption with minimal changes.
• Delivered a tool that is both technically advanced and socially impactful.

What we learned

• How to orchestrate agentic workflows using LangGraph.
• Hands-on experience with TiDB Cloud, including real-time querying, vector search, and schema design.
• How to make semantic vector search truly useful with deduplication, metadata filters, and intelligent chunking.
• The importance of UX and explainability in high-stakes domains like financial complaints.

What's next for ComplaintsCompass

• Expand the dataset to include Reddit complaints, BBB reports, app store reviews, etc.
• Improve risk engine with supervised learning from complaint outcomes (e.g., resolved vs escalated).
• Integrate multi-turn conversation using memory + follow-up agents to handle longer guidance sessions.
• Dashboard for companies and regulators to monitor complaint trends, high-risk companies, and regional breakdowns.
• Public API and open deployment so others can plug ClariRisk into their own apps, bots, or services.

We believe ComplaintsCompass has the potential to significantly improve how consumer complaints are handled, monitored, and resolved—making the process smarter, fairer, and more transparent for everyone involved.

Built With

• nextjs
• python
• tidb