Two engines. No dead time.
A custom on-device speech model with automatic Apple Speech failover while the resumable, SHA-256-verified model downloads.
CS @ BITS Pilani · Builder · Researcher
Sourabhbuildscool stuff.
I turn half-serious “what ifs” into real software — usually with AI, occasionally with robots, and always with too many browser tabs.
Scroll to inspectBuilt inPilani29.9° N / 75.6° E
ON-DEVICE AI✦SWIFT✦1900 CODEFORCES @ 14✦H100 PIPELINES✦REINFORCEMENT LEARNING✦ON-DEVICE AI✦SWIFT✦1900 CODEFORCES @ 14✦H100 PIPELINES✦REINFORCEMENT LEARNING✦
00Short version
I don't collect titles.
I collect hard problems.
First-year Computer Science student at BITS Pilani. Started building Python games in grade six, hit 1900 on Codeforces at fourteen, then wandered into machine learning, product, systems, and native macOS engineering.
The thread through all of it: learn fast, ship the thing, then make the ugly edge cases behave.
01Flagship build · 2026
PRIVATE. LOCAL. ACTUALLY USEFUL.
Air
Scribe.
Hold Control. Speak. Release. Polished text appears wherever your cursor is — with no account, no server, and no audio uploads.
POSTCHATEMAILGENERAL
“make this sound less like it was written at 3am”
POLISHED OUTPUT0.84s
Make this clearer, sharper, and ready to publish.
⌃
HOLD TO TALK
700MBverified model
0audio uploads
The LLM doesn't get the last word.
A diff-based validation layer rejects rewrites or translations that lose names, numbers, or meaning. Fidelity wins over flair.
Four modes. Any text field.
Post, chat, email, and general modes; per-app rules; local history; meetings; and context-aware cleanup — wrapped in a native macOS menu-bar app.
SHIP NOTE / JUL 24, 2026
“this whole post was written using AirScribe.”
25 reactions7 commentsMIT licensed
02Other things that escaped localhost
Selected
chaos
Products, models, agents, and robots. Different surfaces. Same urge to make the idea real.
02A
CONSUMER AI / SHIPPED
GiftSenseAI
A gift recommendation product powered by Gemini — validated by people who actually paid for it.
1.8Kcustomers300 paying $5/mo
02B
REINFORCEMENT LEARNING
Othello agent
Behavioral cloning from expert games, then reinforcement learning until random agents stopped being interesting.
99%win rate
02C
MARKET DATA / MACHINE LEARNING
Quadeye challenge
Built a prediction pipeline for an anonymized market variable and climbed near the top of a very crowded leaderboard.
RANK18/ 3,390 submissions
02D
COMPUTER VISION / EDGE AI
Pest detection, but portable.
Fine-tuned three generations of YOLO on the IP102 benchmark, then exported to ONNX for cross-platform deployment.
03Research & field notes
Currently
in the lab.
NYU Stern School of Business
Machine Learning Research Assistant
Pricing research, web-scale data pipelines, BERT and modern embedding models running on H100 GPUs.
Prahlada OpenAI Research
Founding member · SystemML
Tinygrad, vLLM, AMD ROCm, smaller Docker images, and the stubborn problem of laptop-cluster inference.
Kalipatnapu AR & VR Lab
AI Researcher
Gaze and head-pose models for ADHD screening; hyper-realistic shadow generation; TFLite and ONNX optimization.
Team Robocon
Robotics & AI Engineer
Built and trained in-house robots, shipped edge vision models, and helped run RoboWars.
04Receipts, not adjectives
JEE Main
at age 14
score
followers
CGPA
ATF nationwide
Also: JEE Advanced AIR 2803, IOQM qualified, Aryabhata Ganit Challenge winner, STSE rank 2.
THE TOOLBOX, CURRENTLY
PYTHON+SWIFT+PYTORCH+REACT+C+++ONNX+LLMs+ROCm+SQL+COMPUTER VISION