# Arsh Javed — Full Technical Context & Engineering Dossier

> Comprehensive architecture documents, research paper methodologies, systems design decisions, and background for LLM and agentic crawlers.

- 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

---

## 1. Biography & Professional Summary

Arsh Javed is a Software Engineer and Computer Science undergraduate at Graphic Era University (Class of 2027) based in Dehradun, Uttarakhand, India. His work spans distributed systems engineering, high-concurrency event-driven backends, federated deep learning, and multi-modal machine learning. 

He is an accepted first-author researcher at peer-reviewed IEEE and international conferences, a hackathon champion (1st Place at IIT Guwahati Encode 2025; Top 5 Finalist out of 35,000+ at Google GenAI Exchange Hackathon), and the sole creator/maintainer of Emilia Bot, serving over 800,000 active users across 5,000+ communities.

---

## 2. Research & Academic Publications (First-Author)

### Paper 1: Reconstructing Global Models from Federated Client Weights
- **Venue**: IC2E3 2026 · NIT Uttarakhand (International Conference on Computing, Communication and Energy Systems)
- **Status**: Accepted, Presented, First-Author
- **Problem Statement**: In non-IID federated learning environments, standard aggregation algorithms (such as McMahan's FedAvg) suffer from severe client drift and weight divergence when distributed client datasets exhibit extreme label skew and statistical heterogeneity.
- **Methodology**: Researched and evaluated a learned aggregation mechanism combining four non-IID MentalBERT transformer classifiers. Client weight deltas are reconstructed and aggregated through a parameterized regression mapping rather than naive coordinate-wise averaging.
- **Empirical Results**:
  - Test Set: 4,432 benchmark samples.
  - Achieved **93.50% test accuracy** and **R² = 0.953**, significantly outperforming standard Federated Averaging (FedAvg at **83.98%**) and approaching the upper-bound centralized training baseline (**94.04%**).
- **Artifacts**:
  - Code: https://github.com/ArshCypherZ/MentalBert
  - Pretrained Model: https://huggingface.co/arshjaved/Final-FL
  - Training Notebooks: https://www.kaggle.com/code/arshjaved/novel
  - Certificate: https://files.arshjaved.in/u/bvRCSx.pdf

### Paper 2: Framing, Sentiment, and Agenda-Setting Bias in News Coverage
- **Venue**: IEEE INDISCON 2026 · MNIT Jaipur (Flagship IEEE India Council International Conference)
- **Status**: Accepted, Presented, First-Author
- **Problem Statement**: Quantifying subtle ideological and political framing shifts across heterogeneous news outlets at scale requires multi-layered natural language processing combined with rigorous statistical validation rather than simple dictionary-based sentiment scoring.
- **Methodology**: Analyzed a corpus of **58,289 articles** across 5 national news organizations over longitudinal publication cycles. Built an automated computational pipeline integrating RoBERTa (fine-grained sentiment polarity), BART (abstractive framing distillation), spaCy (entity co-occurrence networks), and BERTopic (unsupervised topic modeling). Applied non-parametric statistical hypothesis testing (Mann-Whitney U, Kruskal-Wallis) to validate systemic framing differences.
- **Artifacts**:
  - Dataset & Code: https://www.kaggle.com/datasets/arshjaved/political-bias-analysis

---

## 3. Production Systems & Engineering Architecture

### 1. Leadly (http://leadly.live)
- **Category**: Outbound Lead Intelligence & Social Selling Platform (SaaS)
- **Core Stack**: Next.js, Express, BullMQ, Redis, Prisma, Docker, Dodo Payments, Gemini 1.5
- **Architecture**:
  - Continuously monitors high-intent Reddit subreddits and query clusters via Reddit's streaming API.
  - Asynchronous BullMQ worker queues decouple rate-limited upstream platform ingestion from token-metered downstream LLM generation.
  - Dual-pass Gemini classification pipeline: Pass 1 filters commercial buying intent from noise (scoring 0–100); Pass 2 drafts organic, subreddit-compliant replies avoiding spam detection.
  - Redis distributed locking prevents duplicate message processing across horizontally scaled worker instances.
  - Dodo Payments integration with automated tier provisioning and token quotas.

### 2. Emilia Bot (https://t.me/Elf_Robot)
- **Category**: High-Concurrency Distributed Automation Platform for Telegram
- **Scale**: 800,000+ active users, 5,000+ Telegram supergroups, 73+ loadable plugins
- **Core Stack**: Python (asyncio), Pyrogram, Telethon, Redis, MongoDB, Groq AI LPUs
- **Architecture**:
  - Employs a dual MTProto client driver (Pyrogram + Telethon) on a custom asyncio event loop to multiplex network sockets and handle burst message traffic exceeding 10,000 messages/min.
  - Multi-tier write-behind cache with Redis: frequent group permission checks, ban lists, and anti-flood configurations are served in under 1ms from RAM, reducing MongoDB IOPS by 95%.
  - Distributed cross-chat federation protocol: allows supergroup networks to synchronize ban lists instantly across thousands of chats during coordinated spam attacks.

### 3. Dialmate Live Voice Agent (https://github.com/arshcypherz/dialmate-backend)
- **Award**: 1st Place Winner at IIT Guwahati Encode 2025 Hackathon
- **Category**: Real-Time Full-Duplex Conversational Voice AI
- **Core Stack**: FastAPI, Pipecat, Silero VAD, Twilio WebSockets, Deepgram, Cartesia TTS
- **Architecture**:
  - Achieves sub-500ms voice round-trip latency (caller speech end to AI voice output).
  - Integrates Silero Voice Activity Detection (VAD) for natural interruption handling: callers can speak over the AI, instantly pausing audio synthesis and recalculating LLM context.
  - Twilio SIP trunking with bi-directional streaming WebSockets handles packet jitter and frame sequencing.
  - Live supervisor dashboard with real-time sentiment analytics, transcript streaming, and human operator intervention.

### 4. Exo-Checkmate (https://github.com/arshcypherz/Exo-Checkmate)
- **Category**: Multi-Modal Deep Learning for Exoplanet Transit Verification
- **Core Stack**: PyTorch, ONNX Runtime, scikit-learn, NASA TESS Photometric Data
- **Architecture**:
  - Hybrid neural architecture combining:
    1. 1D-CNN: extracts local transit ingress/egress morphology from folded photometric light curves.
    2. Bi-LSTM with Temporal Attention: captures long-range orbital periodicity and highlights genuine planetary dips over stellar flares.
    3. Physics-Informed MLP: injects stellar metadata (radius, surface gravity log(g), effective temperature).
  - Optimized with custom FocalLoss to overcome extreme 1:100+ positive-to-negative class imbalance.
  - ONNX Runtime export delivers 14x faster inference throughput for high-volume survey datasets.

### 5. ClarityHub (https://github.com/arshcypherz/ClarityHub)
- **Category**: Document-to-Video Engine & Automated Pedagogical Studio
- **Core Stack**: Next.js, FastAPI, Celery, Redis, Manim, Kokoro TTS, Gemini Flash
- **Architecture**:
  - Transforms technical papers and markdown documentation into narrated 1080p video lessons.
  - Hierarchical semantic decomposition via Gemini parses complex docs into scene graphs and visual metaphors.
  - Sandboxed Python Manim execution engine with automated AST self-healing fixes vector animation syntax errors in real time.
  - Celery distributed worker tasks parallelize video frame rendering and audio alignment.

### 6. DrishtiMind (https://github.com/ArshCypherZ/DrishtiMind)
- **Category**: AI Mental Wellness Toolkit Tailored for Indian Youth
- **Core Stack**: Next.js 14, Tailwind CSS, Framer Motion, Prisma, PostgreSQL (Supabase RLS), Clerk
- **Architecture**:
  - Interactive CBT (Cognitive Behavioral Therapy) reframing coach guiding users through cognitive distortions.
  - Longitudinal mood analytics, streak tracking, and encrypted private journaling with Row-Level Security.

---

## 4. Technical Arsenal & Proficiencies

- **Languages**: Python, TypeScript, JavaScript, SQL, C++, HTML5, CSS3
- **Machine Learning & AI**: PyTorch, Hugging Face Transformers, ONNX Runtime, scikit-learn, Manim, LightGBM, RAG, Multi-Modal Systems, Federated Learning
- **Backend & Distributed Systems**: FastAPI, Node.js, Express, Asyncio, Redis, BullMQ, Celery, WebSockets, WebRTC
- **Databases**: PostgreSQL, MongoDB, Supabase, SQLite, Redis, Prisma ORM
- **Cloud & DevOps**: Docker, Linux (Arch / Ubuntu), Git, Cloudflare Pages, Oracle Cloud Infrastructure (OCI Generative AI Certified)

---

## 5. Contact & Links

- **Website**: https://arshjaved.in/
- **GitHub**: https://github.com/ArshCypherZ
- **LinkedIn**: https://www.linkedin.com/in/arshjaved/
- **Telegram**: https://t.me/halalneko
- **Hugging Face**: https://huggingface.co/ArshCypherZ
- **Kaggle**: https://www.kaggle.com/arshjaved