I work on agent systems under hard constraints, mostly on-device inference budgets and the safety layer that decides what an agent is allowed to do. CS and Artificial Intelligence double major at Purdue, graduating May 2027. Most recently at Samsung Research America on Bixby Edge AI.

karthikthyagarajan.com | LinkedIn | arXiv

Samsung Research America. Two separate projects. Both are written up in my notes.

On Aurora, Samsung's general agent orchestrator, which starts on XR glasses and generalizes to phone and any other Android device, I shipped the permission system, tools, skills, modes, A2UI, MCP integration, dynamic tool and skill discovery, prompt caching, the XR integrations, and computer-use style phone use. I also helped consolidate several competing prototypes into one runtime.

On the on-device signal router for ambient AI, I built the router, the task scopes, and the eval harness. It runs a small local model to decide which incoming signals justify waking a frontier one, suppressing 95% of them at 90% intent-routing accuracy.

PARM is a causal benchmark for output-conditioned memory in AI agents, and my own research. The failure it targets is an agent that takes an ordinary request and then meets a decisive cue inside a tool result, somewhere prompt-time retrieval never looked.

google-tools-mcp puts 169 tools across Drive, Docs, Sheets, Slides, Gmail, Calendar, and Forms behind a single OAuth flow. Agents have to read a document before they can edit it, authentication happens lazily rather than up front, and one install can hold several account profiles. MIT licensed, published to npm, and written up here.

gantry is a coding-agent skill that forces a design gate before any code gets written. You author pseudocode and the agent resolves references and surfaces edge cases inline. The implementation step fails closed until you explicitly approve it. What gets recorded is the decision and its reasons, so the code stays comprehensible months later. MIT.

More on how I think about agents is in this note.

Peraton Labs. Reinforcement learning for IoT malware detection, inside a classified defense research environment. The agent gets low-level network traversal actions and full state observability, which lets it reach device environments a brute-force scan misses. Exploration latency dropped 35% and detection coverage rose 25% across 500K+ daily device events. A heterogeneous graph architecture over device-to-device communication patterns cut policy convergence time 40%.

Memories.ai. A video memory framework for hours-long video that vision language models cannot ingest directly. Keyframe extraction, dynamic segmentation, CLIP indexing, and YOLO and SAM detection index 10K+ streams, and multimodal search runs over transcript, visual, and entity embeddings with graph traversal across scenes, objects, and named people. I published pymavi, the company's Python SDK, to PyPI.

IDEAS Lab, Purdue. Real-time SLAM and neural radiance fields. RGB-D streams fuse with Kalman filtering into a multi-GPU reconstruction pipeline, and the system produces navigation signals from maps the robot has not finished building.

Naval Research Laboratory. Deep learning surrogates for underwater acoustics. Sound speed profiles are encoded as images, and an image-to-image translation model predicts transmission loss 20% more accurately than the physics simulator it replaced, on far less compute.

Photonic Implementation of Quantum Key Distribution, arXiv:2509.04389. Polarization-based QKD prototype with near-infrared lasers and a 95% photon detection rate over a 5m fiber channel, with QBER analysis automated across 10K+ oscilloscope samples. Notes, code.

Repple is a competitive fitness app live on the iOS App Store with 200+ active users. I co-founded it and built the backend in FastAPI, Postgres, and Supabase. It runs real-time synchronized multiplayer workout sessions and an ELO matchmaker that pairs users by fitness level, with cron-driven point aggregation behind the weekly team matchups. Notes.

buildpurdue is a founder-led community for student builders at Purdue, built on the premise that a student founder two months ahead of you is more useful than a curriculum. I'm Co-President and co-founder. I shaped the operating framework and led the build of the internal platform in Next.js and Supabase. Notes.

Veritas pairs World ID iris biometrics with GPT-4o response scoring to keep fabricated participants out of clinical trials. Won Best Proof-of-Human Application at Catapult 2026. Notes, live.

Caladrius is a privacy-first multi-agent triage assistant built with input from practicing physicians, at 85% triage accuracy. 2nd place in HackGT 12's social impact track. Notes.

personalsite is the source for karthikthyagarajan.com, including the retrieval-augmented chatbot that answers questions about my work from a hand-curated corpus.

Long-form essays at karthikthyagarajan.com/blog, mostly on what AI does to work, income, and civilization.

On the Future of Work with AI argues we will not get AGI on the timeline the field assumes, defines what would actually qualify, and describes what human work becomes without it.

Stability in the Age of AI is about why societal stability rests on stable income, and what could take its place.

The AI Company I'd Love to Create is a company I would want to build, written with the holes poked in it.

My Trip to Silicon Valley is a trip report on automation, quality control, AR/XR, and robotics.

I also make videos about how people build things.