Public digital services in developing economies have become central to citizens’ access to government functions, yet frequent deployment failures undermine trust and service delivery. This study investigated whether multilingual citizen feedback could serve as an early-warning signal for deployment failures in resource-constrained public sector digital services. A bimodal failure detection framework was developed combining English system logs with citizen feedback in Akan, Ewe and Ga languages from a national citizen service portal in Ghana handling over 50,000 monthly users. Rather than building language-specific classifiers requiring annotated training data, the study leveraged instruction-tuned large language models as general-purpose interpreters using few-shot prompting. Analysis of 18 months of deployment logs and 1,217 citizen feedback messages revealed that a fusion model combining XGBoost for log analysis and prompt-engineered Qwen 2.5-3B-Instruct for feedback interpretation achieved 82 per cent recall and 76 per cent precision, representing a 21 percentage point improvement over log-only approaches. Citizen feedback preceded technical alerts in 32 per cent of failure cases, with an average detection advantage of 2.3 hours, which increased to 3.8 hours during peak usage periods. This study provides empirical evidence that instruction-tuned large language models can extract operationally useful failure signals from low-resource languages and improve post-deployment early warning when fused with conventional log-based monitoring. The findings imply that public institutions in multilingual developing contexts can adopt LLM-based monitoring without requiring annotated datasets or language-specific processing pipelines.