Pradeep Rajasekar

IdeasLab Formosa, Taipei, Taiwan

• Led R&D on the 2D-to-3D human pose lifting system behind XView AI, a markerless golf-swing analysis app: reduced mean per-joint 3D error 8 cm → 3 cm (pelvis-relative) through temporal consistency modelling, motion-aligned lifting and spatial refinement, with +30% reconstruction fidelity under occlusion and fast motion.
• Authored the technical proposal extending that work to a new sport — arguing how far a golf-tuned lifting stack generalises to a different high-speed motion class — and won TAITRA’s “Go Healthy with Taiwan” 2025 award, 1 of 3 winners from 638 proposals across 55 countries; selected to present the technology at Taiwan Expo Europe 2026.
• Shipped the full analysis pipeline on-device on Apple ARM silicon via Core ML — no cloud round-trip, no footage leaving the handset — in an iOS app used by PGA Tour professionals.
• Extended the pose model to club tracking without regressing the general body model: froze the backbone and every body-keypoint channel and trained through a parallel adapter, leaving the deployed body pathway bit-identical to the base model (verified by weight diff) while the adapter improved club accuracy.
• Diagnosed and corrected calibration and reconstruction faults across the multi-camera stack — 4.8× lower reconstruction noise and −73% event-detection timing error — validated against published anthropometric ratios across every reconstructed session.
• Rebuilt the evaluation behind model releases: recording-level splits that removed train/test leakage, coverage gates that fail incomplete or stale runs, and acceptance thresholds fixed before any result is seen.
• Architected an LLM coaching system in which a deterministic rule engine computes every number and the model writes the explanation, in 8 languages. Rules that fired in every session were reviewed as likely faults, and several were removed from scoring.

President Information Corp (統一資訊), Taipei, Taiwan

• Led the vision architecture for a real-time planogram compliance system — dense detection for localisation, a fine-tuned triplet-loss embedding space for identity — so adding a product costs vectors instead of a retraining run. Deployed across 7,000+ 7-ELEVEN stores in Taiwan; the system’s results (99.23% / 98.93% precision / recall on shelf detection) are documented in Ou et al., Scientific Reports (2025).
• Presented the methodology as a technical poster at NVIDIA GTC 2025.
• Deployed occlusion-robust handheld product recognition into Taiwan’s 8th unmanned 7-Eleven store, a +30% recognition gain under motion blur and arbitrary orientation.
• Built production inference on NVIDIA Metropolis microservices, TAO Toolkit, DeepStream and TensorRT across cloud and edge, in collaboration with NVIDIA.
• Built an LLM-assisted named-entity-recognition pipeline on GCP Document AI, and a predictive model over 5M+ data points at 95% confidence informing marketing and inventory decisions.

AIBS Software Solutions, Coimbatore, India

• Built full-stack ERP systems for manufacturing workflows; engineered inventory and tax tracking that resolved 95% of reported discrepancies.