Soham Chakraborty

Smart glasses that measure eye pressure with ultrasound. I wrote the firmware, designed the board, and trained the models. Now in human trials and heading into FDA pre-submission.

Models that detect Alzheimer's and motor speech disorders from a voice recording. Fully funded on site fellowship. First author on the paper, 1,500 patients in the pipeline.

Designed a space based AI data center: 105K GPUs, thermal and radiation handling, ten year revenue model. Presented it to the New Glenn strategy team.

Haptic gloves and vests paired with a LiDAR headset so blind users can navigate. On edge vision at 96% mAP and 30ms. Designed the two layer board with a team of ten.

Nanoparticle drug delivery for subglottic stenosis, a fibrosis that narrows the airway in newborns. Ran the synthesis, cultures, and in vivo testing.

Analyzed electromagnetic radiation from galaxy clusters using observatory data. Built deep learning classifiers for Faraday spectra that beat the baseline by 14%.

Glasses that measure eye pressure with ultrasound so glaucoma patients can check at home instead of waiting six months. 92% accuracy, 95% cheaper than a clinic tonometer. Patented; clinical study on n=50 patients upcoming.

CNNs that catch Alzheimer's and track its progression from a voice recording. This paper got me the Mayo Clinic fellowship, where I scaled it to a 1,500 patient trial at 88% accuracy.

A vest that lets blind users navigate by feel. I built the driver circuit and a CNN inference stack that runs entirely on an ESP32, so it works with no phone and no internet.

A 0.45 kg drone with a deployable spool electrode that lands, samples soil and water, and flies on. Covers a farm in one pass. 98.4% accuracy on crop stress, 500ms telemetry.

An agentic job search tool that finds roles, tailors each application, and submits it. The agents learn from which submissions convert. 74 users so far.

Collagen binding nanoparticles that release drugs slowly at the site of subglottic stenosis, an airway fibrosis in newborns. Synthesized, cultured, and tested in six mouse models.

A pipeline that takes clinical imaging from raw data to a deployable model: synthesis, training, 4-bit QLoRA finetuning of 7B params, and benchmarking across nine datasets.

A tool that turns NASA stellar luminosity and radiation data into sound so researchers can hear patterns they would otherwise have to plot. 3rd of 1,300 at NASA Space Apps.