Agents & RAGRAG + expert memoryTelecom · home networking
WiFi log intelligence with RAG
Turns raw router logs into plain-language root causes — “Why is my internet slow?” answered with evidence and a next action.
View case study- Home
- Work
Selected work
Work that made it out of the notebook.
15 projects delivered by the Saerosense founding team for companies, hospitals and research programs — each with its workflow, the techniques behind it, and the results it produced.
Agents & RAGRAG + expert memoryAgents & RAG
한국어RAG chatbot · voice channel
Insurance · South Korea
Korean insurance chatbot & voice assistant
A Korean-language RAG chatbot grounded in an insurer’s own documents, benchmarked against ChatGPT, with a phone channel built on speech-to-text and text-to-speech.
View case study
Agents & RAG8 agents · local LLMFinance · investment research
Multi-agent equity research
Eight specialist agents turn 50+ market metrics into a scored BUY/HOLD/SELL research report — on a local LLM, so data stays in-house.
View case study
Computer vision5 damage classesInfrastructure inspection · South Korea
Concrete defect segmentation
Pixel-level detection of cracks, spalling, rebar exposure and efflorescence for a GPR & vision scanner platform.
View case study
Computer vision>90% accuracySmart city · traffic & parking
Blurry license plate recognition
A five-stage pipeline that reads low-resolution, motion-blurred CCTV plates with above 90% accuracy.
View case study
Computer vision7 PPE classesIndustrial safety
PPE compliance detection
A YOLO-NAS detector that checks seven types of protective equipment on workers at construction, manufacturing and mining sites.
View case studyComputer vision
94.8%AUC-ROC · genuine vs counterfeit
Retail · brand protection
Luxury logo authentication
A ConvNeXt-Base pipeline that separates genuine from counterfeit luxury-brand logos — up to 94.8% AUC-ROC.
View case study
Computer visionDrivable space + lanesAutonomous driving
Environment perception for self-driving
Drivable-space and lane estimation from semantic segmentation, with RANSAC ground-plane fitting.
View case study
Computer vision<0.2 m errorRobotics · drones · AR/VR
Stereo visual-inertial odometry (MSCKF)
Drift-reduced motion estimation fusing stereo cameras and an IMU — position error under 0.2 m across the evaluated trajectory.
View case study
Computer vision3D U-NetMedical imaging
Brain tumor segmentation in MRI
A 3D U-Net that segments brain tumors from multi-modal MRI volumes, trained with a multi-class soft Dice loss.
View case study
Computer vision3 generative familiesMedical imaging · radiology
Generative AI for knee X-rays
Knee radiograph re-generation with StyleGAN3, Stable Diffusion and Pix2Pix GAN, in partnership with a university hospital in Seoul.
View case studyNetwork & telemetry
93.5%Test accuracy · 230 KB on-device model
Smart home · telecom
On-device WiFi sensing
CNN + BiLSTM + attention recognizes human activity from WiFi signals — 93.5% accuracy in a 230 KB model on a 1 GB ARM device.
View case study
Network & telemetryIEEE TPWRDPower systems · transmission
Auto-reclosing with particle filter & CNN
Detects secondary-arc extinction and permanent faults on transmission lines so breakers reclose only when it is safe. Published in IEEE Transactions on Power Delivery.
View case studyNetwork & telemetry
90.5%Permanent-fault accuracy on unseen lines
Power systems · transmission
Pattern-recognition auto-reclosing with Bi-LSTM
A Bi-LSTM predicts the secondary-arc extinction interval and fault type — 90.5% permanent-fault accuracy on unseen line models. Published in IEEE Access.
View case studyNetwork & telemetry
0.0018RMSE · CatBoost + PCA
Urban infrastructure · climate risk
Flood depth prediction
CatBoost and PCA over multi-resolution simulation rasters predict flood depth — RMSE 0.0018, MAE 0.0002.
View case studyClient names are withheld unless public. If you want references relevant to your industry, ask us directly.
Ask for referencesStart a project
Tell us about your data.
Two or three sentences about the data you have and the outcome you want is enough. We reply within two business days with an honest read on feasibility.