Tideglow Labs is our internal research group. We study how modern AI models can run efficiently, safely and fairly on the phones people already own, and we turn the findings into product features.
Post-training quantization (INT8/INT4), structured pruning and knowledge distillation so vision and language models fit within mobile memory and thermal budgets. We measure latency, energy per inference and accuracy loss on mid-range Android devices, not only flagship hardware.
How should an assistant show uncertainty? When should it ask instead of act? We run moderated usability studies and A/B experiments on explanation styles, undo patterns and progressive disclosure for AI features.
Local-first architectures, on-device personalization and data minimization. We look at federated evaluation and differential privacy for product analytics wherever they are practical.
Speech, text and image models that work for Vietnamese and other under-served languages. Our focus is robust speech recognition, natural TTS and OCR for real-world photos.
Short technical notes from Tideglow Labs that summarize the experiments behind our products. Full reports are available to partners on request.
We compare per-tensor and per-channel INT8 quantization for lightweight segmentation backbones and report the trade-off between mask quality, latency and energy. The study informs our background-removal pipeline.
Computer visionQuantizationAndroidA within-subjects usability study comparing three ways of showing model confidence in a chat-style mobile assistant. Participants preferred short natural-language hedges to numeric scores.
HCIUser studyLLMWe evaluate noise augmentation and lightweight fine-tuning strategies for on-device speech recognition in Vietnamese, using recordings from cafés, streets and vehicles.
SpeechLow-resource languagesA reference architecture that keeps user content on the device, sends only minimal and explicitly consented data to cloud models, and fails gracefully offline.
PrivacyArchitectureiOSA feature only ships when it passes all four gates. Lab accuracy alone isn't enough.
We measure accuracy on held-out and adversarial test sets that reflect real usage, not only public benchmarks.
We record p50/p95 latency, peak memory and energy across a matrix of low-, mid- and high-end devices.
We check for bias, harmful outputs, privacy exposure and failure modes, with documented mitigations.
Staged rollouts with opt-in feedback. We judge success by whether users finish their task, not by clicks.
Students in computer science, HCI and data science can join a 3–6 month research project with a mentor, working toward a publishable technical report.
An internal weekly seminar where we review recent papers on efficient deep learning, multimodal models and AI safety.
We welcome collaboration with universities and research labs on applied mobile AI. We can share anonymized benchmarks and real-device testing.
Researchers, students and partners are welcome to reach out.
Contact the research team →