# Arsh Javed

> Software builder and ML engineer. Probably building something late into the night, petting cats, or making distributed services fast.

Hey, I'm Arsh (Arsh Javed). I'm a Computer Science undergrad at Graphic Era University (Class of 2027) based in Dehradun, Uttarakhand, India. Most of my time goes into training ML models, building high-concurrency Telegram bots, and shipping full-stack systems that handle real production load.

## Quick Highlights

- **IIT Guwahati Hackathon Winner**: 1st Place at Encode 2025 with Dialmate, an ultra-low latency voice agent.
- **Google GenAI Hackathon**: Top 5 Finalist out of 35,000+ participants with PolicySense.
- **2 Accepted First-Author Papers**: Peer-reviewed research accepted and presented at IEEE & international conferences.
- **800,000+ Users**: Created and actively maintain Emilia Bot, a modular automation bot serving over 5,000 Telegram communities.
- **Amazon ML Challenge 2025**: Ranked ~600 out of 80,000 registrations (top 0.75%).

## Research & Publications (First-Author)

- [Reconstructing Global Models from Federated Client Weights](https://files.arshjaved.in/u/bvRCSx.pdf)  
  *Venue*: IC2E3 2026 · NIT Uttarakhand  
  *What it does*: Researched learned aggregation of 4 non-IID MentalBERT classifiers in federated learning. Hit 93.50% accuracy on a 4,432-sample test set (beating standard FedAvg's 83.98% and coming close to centralized training at 94.04%) with R² = 0.953.  
  *Links*: [Code on GitHub](https://github.com/ArshCypherZ/MentalBert) · [Model on HuggingFace](https://huggingface.co/arshjaved/Final-FL) · [Training on Kaggle](https://www.kaggle.com/code/arshjaved/novel) · [Certificate](https://files.arshjaved.in/u/bvRCSx.pdf)

- [Framing, Sentiment, and Agenda-Setting Bias in News Coverage](https://www.kaggle.com/datasets/arshjaved/political-bias-analysis)  
  *Venue*: IEEE INDISCON 2026 · MNIT Jaipur  
  *What it does*: Large-scale computational framing analysis across 58,289 articles from 5 major news outlets. Combined RoBERTa, BART, spaCy, and BERTopic topic modeling with nonparametric statistical validation to quantify media framing shifts.  
  *Links*: [Dataset + Code on Kaggle](https://www.kaggle.com/datasets/arshjaved/political-bias-analysis)

## Projects & Systems I've Built

- [Leadly](http://leadly.live)  
  *Turn Reddit conversations into revenue.*  
  Production SaaS that monitors relevant Reddit threads, scores buyer purchase intent with Gemini AI, and generates authentic, community-compliant replies. Built with Next.js, Express, BullMQ, Redis queue clusters, Prisma, Docker, and Dodo Payments.

- [Emilia Bot](https://t.me/Elf_Robot)  
  *Modular automation bot for Telegram.*  
  Serves 800,000+ active users across 5,000+ groups with 73+ loadable plugins. Uses an asynchronous Python core (Pyrogram + Telethon) with a Redis write-behind cache that cuts MongoDB database load by 95%.  
  *Link*: [GitHub](https://github.com/arshcypherz/Emilia)

- [Dialmate Live Agent](https://github.com/arshcypherz/dialmate-backend)  
  *Sub-500ms full-duplex conversational voice agent.*  
  1st Place winner at IIT Guwahati Encode 2025. Built with Pipecat, FastAPI, Silero VAD (handles real-time caller interruptions), and Twilio WebSockets for natural turn-taking over the phone without awkward delay.

- [Exo-Checkmate](https://github.com/arshcypherz/Exo-Checkmate)  
  *Custom ML architecture for exoplanet transit verification.*  
  Hybrid deep learning pipeline combining a 1D-CNN, Bi-LSTM with temporal attention, and a physics-informed MLP to classify planetary candidates from NASA TESS photometric light curves. Tackles 1:100+ class imbalance using FocalLoss and exports to ONNX Runtime for 14x faster survey throughput.

- [ClarityHub](https://github.com/arshcypherz/ClarityHub)  
  *Doc-to-lesson video studio.*  
  Converts technical papers and documentation into 1080p narrated video lessons. Uses Gemini to break docs into scene graphs, sandboxed Python Manim with self-healing syntax repair for vector math animations, Kokoro neural TTS for voice, and Celery + Redis for rendering.

- [DrishtiMind](https://github.com/ArshCypherZ/DrishtiMind)  
  *AI mental wellness toolkit tailored for Indian youth.*  
  Features an AI CBT perspective coach, longitudinal mood analytics, and private journaling. Built on Next.js 14, Tailwind, Framer Motion, Prisma, PostgreSQL on Supabase with Row-Level Security, and Clerk auth.

## Tech I Work With Daily

- **Languages & Frameworks**: Python, TypeScript, React, Next.js, FastAPI, Node.js, Express, Flask
- **Machine Learning**: PyTorch, Hugging Face Transformers, scikit-learn, ONNX Runtime, Manim, LightGBM
- **Databases & Cache**: Redis, PostgreSQL, MongoDB, Supabase, SQLite, Prisma ORM
- **Infrastructure & Tools**: Docker, Linux, Celery, BullMQ, Git, WebRTC, Twilio, OCI (Oracle Cloud Infrastructure Generative AI Certified)

## Contact & Links

- Website: https://arshjaved.in/
- GitHub: https://github.com/ArshCypherZ
- LinkedIn: https://www.linkedin.com/in/arshjaved/
- Hugging Face: https://huggingface.co/ArshCypherZ
- Kaggle: https://www.kaggle.com/arshjaved
- Telegram: https://t.me/halalneko
- Email: Let's talk via https://arshjaved.in/#contact

## Optional

- [Full Engineering Context](https://arshjaved.in/llms-full.txt): Comprehensive technical architectures, paper abstracts, and project dossiers.