I work the part of AI that has to be True

Pradeep Rajasekar, also known as Ajish Pradeep. AI Research Engineer at IdeasLab Formosa, New Taipei City, Taiwan.

I am an AI Research Engineer at IdeasLab Formosa in Taipei. I take research models through deployment on the hardware and data they meet in practice. At IdeasLab that means a phone that rebuilds a golf swing in 3D from two ordinary cameras, and coaching software whose numbers change how someone moves. Before that, it meant recognising products in an unmanned 7-Eleven store.

The difficult failures are often missing from the validation set. I have seen a pose model improve on paper while introducing a field bias, a reconstruction preserve its proportions at the wrong scale, and a language model explain a false result confidently. I check those outputs against real sessions, geometry and deterministic calculations.

Linear algebra, probability, multi-view geometry and optimisation have let me move between generative models, retail vision, biomechanics and LLM systems without relearning the fundamentals each time.

I also do a lot of the writing: research proposals, conference submissions and mathematical walkthroughs. One proposal won TAITRA’s “Go Healthy with Taiwan” award, one of three winners from 638 entries. At IdeasLab I wrote the team’s rules for working with AI coding tools, and I mentor engineers on using those tools without losing the ability to reproduce their results.

Trajectory

From software developer to AI research engineer, newest first. Open an entry to see what I did there.

2025 —AI Research EngineerIdeasLab Formosa, Taipei, Taiwan
- Led R&D of the 2D-to-3D pose lifting system behind a markerless golf-swing analysis app that runs its whole pipeline on the phone.
- Reduced mean per-joint error 8cm → 3cm, pelvis-relative; +30% reconstruction fidelity under occlusion and fast motion.
- Deployed to Apple ARM silicon via CoreML with no cloud round-trip and no footage leaving the device, used by PGA professionals.
- Designing an LLM coaching system where code does the scoring and every point cites its rule.
- Wrote the team's rules, skills and hooks for AI coding agents; mentor engineers on working with them.
- Technical lead on the proposal that won the TAITRA "Go Healthy Taiwan" award; selected to represent engineering at Taiwan Expo Europe.
2023 — 2025AI EngineerPresident Information Corp (統一資訊), Taipei, Taiwan
- Led a real-time planogram compliance system presented as a technical poster at NVIDIA GTC 2025.
- Combined object detection with fine-tuned embeddings, so new products could be added without retraining.
- Deployed occlusion-robust product recognition to Taiwan's 8th unmanned 7-Eleven store, +30% performance.
- Built NVIDIA Metropolis microservices, TAO, DeepStream and TensorRT pipelines across cloud and edge.
- LLM-assisted entity extraction on GCP Document AI; a predictive model over 5M+ data points at 95% confidence.
2021 — 2023MSc, Electrical Engineering & Computer ScienceNational Taipei University of Technology, Taipei, Taiwan · GPA 3.8/4.0
- Thesis: Content and Spatial Aware Generative Model for Inpainting.
- GAN architectures, attention mechanisms and the mathematical foundations of deep learning.
2017 — 2021Software DeveloperAIBS Software Solutions, Coimbatore, India
- Custom ERP systems for manufacturing workflows; inventory and tax tracking resolving 95% of reported discrepancies.
- Full-stack development of those systems, from the database to the screen.
2011 — 2014BSc, Information TechnologySri Ramakrishna Mission Vidyalaya College, Coimbatore, India · GPA 7.9/10
- Foundations in computer science, algorithms and systems design.

Areas of work

Mathematical foundations: Linear algebra, probabilistic modelling, multi-view geometry, optimisation and metric learning. These carry over from one field to the next, and every case study on this site depends on at least one of them.
Computer vision: 2D/3D pose estimation and lifting, dense object detection, open-set recognition by embedding retrieval, 3D reconstruction, kinematic and biomechanical analysis, temporal consistency under occlusion and fast motion.
LLM systems: Fine-tuning on domain data, tool-using workflows, directly addressable rule bases, retrieval pipelines, and evaluations for systems where a wrong answer is expensive.
Deployment & inference: CoreML and ARM edge deployment, TensorRT, DeepStream, quantisation and compression, cloud-native microservices, inference-cost optimisation.
Research leadership: Technical proposals and conference submissions, AI coding standards for engineering teams, mentoring, and working closely with domain experts.

What I’m open to

Research collaborations, hard applied problems in 3D vision or on-device AI, and conversations about either. I am based in Taiwan and open to relocating.