Deployed · v6.16KTH · Forte
Aktivitetshjälpen — activity assistant for the pre-frail
Qwen3-8B agent · KB-Whisper · Piper · hybrid semantic search · FastAPI
A bilingual agentic LLM assistant helping pre-frail older adults find local activities. Tool-calling agent over Qwen3-8B with hybrid keyword-boosted semantic search (nomic-embed-text), honesty guards for weak matches, 11-table data model scraping 26 municipalities + SPF Seniorerna daily, CPU-only speech pipeline (KB-Whisper int8, 4–7s latency), WCAG AA accessible UI, consent-gated family sharing, and admin analytics dashboard.
Status● LIVE · 26 sources · sv/en
Challenge · 3rd of 5Team Hedgehog
VisGrid drone classification
Drone Defense Hackathon · KTH Center for Total Defence
Classify tracked objects as BIRD / DRONE / IRRELEVANT from multi-camera image sequences. Lightweight motion & appearance features feeding a tuned tabular ensemble (ExtraTrees, XGBoost, LightGBM) — optimised for macro-F1 on a small dataset, no GPU.
Macro-F10.9458
Project · clinical NLPKTH · w/ Yuxin Du
Fine-tuning NanoChat on MIMIC-IV discharge notes
transformer LM · domain adaptation · honest eval
Domain-adapted a transformer language model to clinical discharge summaries, then built a QA + hallucination evaluation that quantifies the gap between better language modelling and reliable clinical question answering.
Perplexity34.4 → 9.0
Paper + codeNCVPRIPG 2025
HybridMorph — MR image registration
semi-supervised · spatial transformers · TensorFlow
Aligns medical MR images from hybrid real + synthetic data sources, using spatial transformers to cut the amount of real training data needed for accurate registration.
Accuracy+9.2% limited data
Paper + thesisIIST · in prep
Temperature fields from laser absorption + Gaussian processes
LAS · multi-output GPR · Boltzmann parameterisation
Reconstructs non-uniform temperature profiles inside a combustor from a single molecular absorption transition, using multi-output Gaussian-process regression and testing robustness to spectral noise against experimental data.
RMSE (sim)19.9 · ±20 K exp
Project · meta-learningFS-DSC 2024
MAML vs. FOMAML on few-shot classification
Omniglot · learn2learn · PyTorch
Compared Model-Agnostic Meta-Learning against its first-order approximation on few-shot character recognition, testing when dropping the second-order (Hessian) term is safe — and quantifying the accuracy/speed trade-off.
FOMAML acc0.970 · 5% faster
Research · remote sensingNRSC · ISRO
Complex-valued U-Net for PolSAR segmentation
Pauli decomposition · complex CNN · raw data
Segments raw Polarimetric SAR imagery directly in the complex domain via the Pauli representation, avoiding the domain shift introduced by converting radar data to real-valued inputs.
InputRAW POLSAR · CV-CNN
Study · psychophysiologyReykjavík · NeurotechEU
HRV under cognitive vs. social-evaluative stress
Polar H10 · time + frequency domain · effect sizes
A within-subject HRV experiment testing whether social observation deepens stress. Found parasympathetic engagement, not suppression — a clean case study in reading physiological signals honestly against a directional hypothesis.
Signal1,230 IBIs · |d|>0.8
Earlier research also includes a CNN for wheat-grain damage detection at IIT Delhi (99.82% accuracy), an ESP32 construction-safety edge device (98.3%), FFT + wavelet analysis of combustion instability, and a complex-network study of the OPEC crude-oil trade — see publications below.
