Fact-checking with AI and human oversight.

The engineering challenge

Fact-checking requires several different kinds of work: identifying a claim, finding evidence, interpreting what it supports and communicating a conclusion. A usable product needs to coordinate those steps while giving people a way to review the result and add context.

Rajkumar built the full application and orchestrated LangChain and LangGraph pipelines from ingestion through conclusion reporting. The workflow included claim identification, evidence summarization and report generation, with prompts managed and evaluated using PromptLayer and Vellum.

The retrieval layer used vector embeddings and stores including Qdrant and PgVector to match queries to evidence chunks. This connected the analysis to retrieved material, rather than treating the language model’s answer as sufficient evidence on its own.

Human-in-the-loop steps paired automated analysis with expert review. The product’s multimodal approach covered information from formats such as images, video and audio, with human judgment providing contextual checks on the analysis and reporting.

LiteLLM connected models from multiple providers. FastAPI services exposed querying and evidence retrieval capabilities to downstream applications. Containerized workloads ran on Google Kubernetes Engine, with autoscaling and release pipelines supporting production delivery.

What was delivered

Factso was built, launched and acquired. Rajkumar subsequently integrated it into Verideck during a separate consulting engagement. This case study covers Factso’s independent development before the acquisition.

What this experience brings to your project