In this paper, we present the technical report on the CL4Health 2026 Shared Task on Case Report Form (CRF) filling for our team Innov8rs. The paper explains the complete development of our system for the CL4Health 2026 Shared Task. We describe every phase of our system – from initial catastrophic failures with small models producing over 4,800 false positives, through prompt engineering breakthroughs, to our final multi-LLM ensemble combining Gemini 2.5 Flash and Llama 3.3 70B with dynamic TF-IDF-based few-shot retrieval. The main contribution of this work is a data-driven precision filter that suppresses predictions for CRF items with historically high false-positive rates. This single intervention reduced false positives from 816 to 171 on the English development set, boosting macro-F1 from 0.541 to 0.703. We document the engineering challenges of multi-API-key rotation across 11 Google API keys and 2 Groq keys, the design of four distinct ensemble strategies, and the critical analysis of why development-calibrated filters suffered from distribution shift on test data (final test F1: 0.47).

Smart devices are often deployed in some edge-devices, which require quality solutions in limited amount of memory usage. In most of the user-interaction based smart devices, coreference resolution is often required. Keeping this in view, we have developed a fast and lightweight coreference resolution model which meets the minimum memory requirement and converges faster. In order to generate the embeddings for solving the task of coreference resolution, DistilBERT, a light weight BERT module is utilized. DistilBERT consumes less memory (only 60% of memory in comparison to BERT-based heavy model) and it is suitable for deployment in edge devices. DistilBERT embedding helps in 60% faster convergence with an accuracy compromise of 2.59%, and 6.49% with respect to its base model and current state-of-the-art, respectively.

Automatic prediction of personality traits has many real-life applications, e.g., in forensics, recommender systems, personalized services etc.. In this work, we have proposed a solution framework for solving the problem of predicting the personality traits of a user from videos. Ambient, facial and the audio features are extracted from the video of the user. These features are used for the final output prediction. The visual and audio modalities are combined in two different ways: averaging of predictions obtained from the individual modalities, and concatenation of features in multi-modal setting. The dataset released in Chalearn-16 is used for evaluating the performance of the system. Experimental results illustrate that it is possible to obtain better performance with a hand full of images, rather than using all the images present in the video