B.Tech, Electronics and Communication Engineering
Manipal Institute of Technology, MAHE. Expected 2029.
Drag it, press W or , or scroll.
Drag it up, or swipe up anywhere.
ECE undergraduate, Manipal Institute of Technology (2025–2029).
I work on the physical layer: channel coding, modulation and link simulation, and on the software and embedded systems built around it.
Seeking Summer 2027 research internships in wireless communication and embedded systems.
Push the lever forward to rideDrag it up, or swipe up anywhere.
Simulated Kodak images over AWGN at a 1/12 bandwidth ratio using JPEG, JPEG 2000 and HEVC source coding with 5G NR LDPC (n = 6144, QPSK to 256-QAM, 12 MCS levels) in NVIDIA Sionna. Benchmarked the links against Shannon-capacity and finite-blocklength (normal approximation) bounds from 0 to 20 dB SNR, as baselines for Deep JSCC.
Adaptive MCS shows a sharp cliff under a 1 dB SNR-estimation error: up to 5.7 dB PSNR loss (HEVC at 14 dB SNR).
Real-time carpool matching for riders on overlapping routes, with fares split by distance travelled.
A Postgres matching function on Supabase pairs riders heading the same direction in real time, with no pre-scheduled rides. Each shared stretch is split per km among the riders on it.
Supabase on PostgreSQL for the matching function, row-level security, and email OTP and magic-link sign-in with deep-link handling. APK builds on GitHub Actions CI.
In development, now in on-device testing.
Keeps video playback synchronized across viewers on Netflix and JioHotstar, with end-to-end encrypted chat and WebRTC voice.
The Manifest V3 extension relays play, pause and seek events over WebSockets, and clients auto-resync on reconnect.
WebSockets for state, WebRTC for voice, and Cloudflare Workers with Durable Objects holding each room's state, so rooms persist across redeploys.
Estimates a video's most- and least-replayed segments. It uses YouTube's official replay data when available, and otherwise a fusion of comment timestamps, captions, audio events and visual cues.
A ridge-regression audio head on 21 YAMNet sound-event features, trained with leave-one-video-out validation on YTMR500. An offline evaluation harness scores estimates against official curves (Spearman, top-5 peak recall) with nested model selection and paired bootstrap confidence intervals.
Two ESP32s, one transmitting and one logging WiFi channel state information (CSI), detect and count punches while shadowboxing in one room, along with punch rate and rounds. Classifying punch type (jab, cross, hook, uppercut) comes next.
B.Tech, Electronics and Communication Engineering
Manipal Institute of Technology, MAHE. Expected 2029.
Research Society, Management Committee, AI and Electronics
IEEE EMBS, Working Committee, Research and Projects, Media and Content
Hostel Representative, Block 9
or drag the tonearm onto the record