MS Computer Science @ Stevens (4.0) · Founder & CEO @ Scholera · ex-adMarketplace ML
I build AI that helps people learn — and I'm still trying to prove it can do that without doing the learning for them.
Right now that means Scholera: Scholars, not scores. An AI-native layer that sits on top of Canvas, Blackboard, Moodle, and Brightspace — RAG-powered course-specific tutoring, living roadmaps, an adaptive quiz studio, and co-pilot grading. It started as my research at Stevens, spun out, and is now live in real classrooms.
Before that I was shipping ML infra at adMarketplace, and before that building research systems at IIT Bombay and the Explainable & Controllable AI Lab. Somewhere in between I published an audio denoiser that beat NVIDIA's baseline.
📘 Fun fact: I'm deeply curious about how technology can improve inner clarity, mental well-being, and lifelong learning — one model at a time.
| $320K | 280+ | 500+ | 4 |
| NJ state AI grant secured | teams we beat for it | students on the platform | NJ institutions in reach |
| +2.3 dB | 96.13% | 5x | 3 |
| SNR over NVIDIA CleanUNet | deepfake detection accuracy | hackathons won | open-source libs contributed to |
| Project | What it does | Stack |
|---|---|---|
| Scholera 🏫 | AI-native LMS layer — course-specific RAG tutoring, adaptive quizzes, co-pilot grading. NJ Plug and Play & NJEADA Fellow. | Next.js RAG LangChain Postgres |
| WaveSplit 🎧 | Multi-stage audio denoiser with SNR-aware adaptive filtering, harmonic-percussive decomposition, and psychoacoustic masking. +2.3 dB SNR / +0.21 PESQ over CleanUNet. 📄 Published | PyTorch Transformers ONNX TensorRT |
| Omniscient Reader 🧠 | Chat platform with persistent memory — 95%+ context retrieval across 5K+ conversation histories, plus a "wingman" that drafts replies in <500ms. | Next.js Supermemory MongoDB WebSocket |
| DeepShield 🛡️ | Multimodal deepfake detection fusing CNNs, RNNs, and BERT embeddings. 96.13% under adversarial conditions. 🏆 Aeravat Hackathon winner | TensorFlow OpenCV scikit-learn |
| Adaptive LMS (research) 📚 | DQN + Bayesian Knowledge Tracing with Elo-based skill estimation. +35% engagement, 91% prediction accuracy. → became Scholera. | PyTorch RL BKT |
| Mumbai Grid Platform ⚡ | 50K+ interconnected components for a 20M-resident power grid. Hours → sub-second queries. | Next.js Firebase MongoDB |
- Agent infrastructure — MCP servers, tool layers, prompt caching, and the tenant-isolation work nobody posts about
- RL for adaptive curricula — DQN + Bayesian Knowledge Tracing, Elo-based skill estimation
- Deep learning for audio — transformer denoisers, psychoacoustic masking, streaming inference under non-stationary noise
💬 Ask me about…
- Why shipping the model is maybe 10% of the work, and the tool layer is the other 90%
- What actually breaks when you put an LLM in front of 500 students
- Going from a research prototype to a funded company in under a year
- Why MSE is probably the wrong metric for whatever you're ranking
- Teaching Distributed Computing labs to 60+ students (RMI, RPC, clock sync) and why it made me a better engineer
🏅 Achievements & community
- 🥇 Winner, 5x Hackathons — first place across five, building under tight deadlines
- 🎓 Teaching Assistant, Distributed Computing @ S.P.I.T. — led labs for 60+ students under Dr. Sudhir Dhage on RMI, RPC, and clock synchronization
- 🌍 Open Source Contributor — performance optimizations to
pandas,scipy, andscikit-learn - 🏛️ NJ Plug and Play & NJEADA Fellow — selected from 280+ teams for New Jersey's $3.34M AI education grant pool
Languages
AI / ML
Backend & Frameworks
Data & Infra
If you're building in AI + education, or just want to argue about eval metrics — my inbox is open.