Instruments
Hardware and measurement: the AEGIS plasma gantry and the pupillometry rig. Things that produce numbers you can argue with.
Signal acquired / profile interface
independent researcher // developer // plasma robotics + measurement
A cinematic operating-system portfolio for medical robotics, pupillometry research, local-first systems, and the ideas currently loading.
[scan] syncing profile signal / independent researcher + developer _
✦ cinematic portfolio dossier
A visual portfolio OS with cinematic proof, resume staging, verified project links, and direct contact paths.
Press play to listen. Playback availability is controlled by Spotify. If the player does not load, use the link above.
profile dossier
Researcher & Software Engineer
I'm a student researcher and engineer. I research theoretical computer science at Carnegie Mellon's School of Computer Science and computational information at MIT CSAIL, lead a team at Algoverse evaluating whether language models hold truth consistency against verified evidence, and build DBD plasma robotics for automated wound sterilization at IntelliScience. I'm also CTO of a stealth AI startup and a software engineer at Drivetrain AI.
What ties them together is a preference for projects that can be falsified. A gantry either homes and reaches the target coordinate or it does not. A pupil response either separates imagery from non-imagery or it does not. I would rather build the instrument than describe the idea.
Languages
Research
Systems
Product
verified proof reel
Autonomous Epidermal and Germicidal Imaging System: a 3-axis gantry that detects open wounds with a Pixy2 camera and applies cold atmospheric plasma over an automatically generated toolpath. Arduino Mega 2560, NEMA steppers through DM556 drivers, limit-switch homing, emergency stop, and IR temperature monitoring.
An objective webcam-based measure of visual mental imagery. Aphantasia affects roughly 1-4% of people and is diagnosed almost entirely by self-report questionnaire, which cannot separate 'I have no image' from 'I have an image I cannot describe'. Currently at Phase 0: the noise-floor spike, deliberately scoped before any imagery claim.
A native macOS local-first assistant for reminders, reusable prompts, and bounded parallel work. SwiftUI with deterministic Swift domain logic and system SQLite. No conversational language model, no cloud backend, no shell agent. Optional name detection uses Apple's on-device speech recognizer. 190 tests passing at the last verified checkpoint.
A multi-agent public-equity research desk built on FastAPI. Each request validates live daily market history, runs a Bull-Bear debate, applies an independent risk haircut, and streams the Portfolio Manager verdict as Server-Sent Events. Deterministic momentum scoring, realized-volatility adjustment, and explicit confidence caps.
A standalone open-source hand-pose model detecting palms and estimating 21 image-space and 21 world-space landmarks with OpenCV DNN. CPU-first reproducible inference, deterministic contract tests, opt-in real-model smoke tests, Apache-2.0 source and weights.
A camera-based obstacle alert prototype for low-vision navigation, built with a team at De Anza Hacks 5.0. Browser frontend over a Node.js/Fastify detection backend, plus an iOS sample using packaged offline object detection with optional LiDAR depth estimation.
portfolio atlas
Same dark red system, different posture: less terminal, more editorial spotlight with projects, resume, and contact staged like a visual dossier.
Hardware and measurement: the AEGIS plasma gantry and the pupillometry rig. Things that produce numbers you can argue with.
Local-first macOS tools and multi-agent infrastructure. Explicit boundaries, deterministic logic, and test suites that actually run.
Computer vision releases with reproducible CPU inference, published weights, and contract tests rather than demo notebooks.
Robotics team, IntelliScience, a car crash that turned into collision calculations, and the independent projects that followed.
Use the dock to open projects, research timeline, and contact.
Each project card states what was built, what was measured, and where the source lives.
Open a repository, read the research poster, or email directly.
resume
Student researcher at Carnegie Mellon SCS and MIT CSAIL, CTO of a stealth AI startup, and software engineer at Drivetrain AI. Research spans theoretical computer science, LLM truth evaluation, and DBD plasma robotics for automated wound sterilization. I prefer projects that force a real measurement, a real control loop, or a real privacy guarantee.
trajectory
Research at the intersection of algorithms, mathematical reasoning, and computation.
Synthesizing technical literature and translating open research questions into structured analytical work.
Leading technical strategy and product development for an early-stage AI company recognized by the a16z Speedrun and StartX ecosystems.
Engineering work on an AI-powered financial planning and decision-support platform.
Directing a four-person team evaluating whether large language models hold truth consistency against verified evidence. Designing adversarial tests and a benchmark for truth adherence, robustness, and judge-model dynamics.
Developing a state-recognized biomedical robotics system for automated wound sterilization using DBD plasma. Coauthoring work extending the 3-axis gantry workflow to a six-axis platform.
Built and integrated FTC robot software and hardware, then supported FRC event operations through field management.
Coursework including Data Structures and Abstractions, alongside independent research in robotics, applied AI, and psychophysics.
open channel
Open to internships, collaborations, and conversations about products that deserve to exist.
ComposeHardware
AEGIS
3-axis CAP gantry, Pixy2 wound detection
Research
Phase 0
pupillometry noise floor established
Systems
ARGUS
190 Swift tests passing, zero cloud backend
Contact
direct
yasht2008@gmail.com
quick faq
3 answersResearch internships, lab collaborations, plasma or pupillometry work, or engineering roles where measurable correctness matters more than demo polish.
Email yasht2008@gmail.com.
Every project card links to its repository. AEGIS, ARGUS, Northstar, Open Hand Model, and Phantasia Meter are all public on GitHub.
Undergraduate Research Technology Conference (IEEE URTC)
Poster presenting the plasma treatment research behind AEGIS: the dermatological application of cold atmospheric plasma, the imaging approach for identifying treatment regions, and the hardware path from concept to a working 3-axis delivery system.
output