Skip to content
View Pratyush426's full-sized avatar
🎯
Focusing
🎯
Focusing

Block or report Pratyush426

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
Pratyush426/README.md

Typing SVG


Β  Β  Β 



"Bridging theoretical machine learning with production-grade distributed architectures."


⚑ About Me

const pratyush = {
  currentRole   : "Research Intern @ CAIR, DRDO",
  education     : "B.E. in Electronics & Communication Engineering @ MSRIT Bengaluru (8.53 CGPA)",
  coreDomains   : ["Machine Learning", "Natural Language Processing", "Signal Processing", "Distributed Systems"],
  algorithmic   : "LeetCode 1755 Rating (400+ Solved Β· Top 10%)",
  researchFocus : [
    "Contextual multi-armed bandits & emergent multi-agent decision systems",
    "Cross-modal synchronization fusion for synthetic media & deepfake detection",
    "High-throughput asynchronous microservices & non-blocking inference cascades"
  ],
  engineering   : ["Async by default", "Statistical rigor", "Type-safe end-to-end"],
  availability  : "Open for long-term Research Fellow / Intern appointments (6–10 months) & AI Systems roles",
  location      : "Bengaluru, Karnataka, India"
};

πŸ”¬ Flagship Research & Core Systems

🐝 ClawSwarm AI: Adaptive Multi-Agent Orchestration

Independent Research Framework

Live Demo Source

Eliminates static, manually hardcoded roles in multi-agent frameworks (CrewAI, AutoGen). Agents begin dynamically identical and evolve emergent specializations through empirical performance feedback.

πŸ”¬ Innovations & Benchmarks:

  • Dynamic Routing: Formulated as a contextual multi-armed bandit with 6D Exponential Moving Average (EMA) skill vectors and entropy-decay $\varepsilon$-greedy exploration.
  • Empirical Validation: 63% cumulative regret reduction ($p < 0.001$, paired Wilcoxon signed-rank test) across 50-seed simulation benchmarks with cross-model LLM-as-judge evaluation.
  • Fault-Tolerant Consensus: 2-second peer heartbeat leader election with live D3.js force-directed topology.

Stack: Python FastAPI Redis Claude 3.5 Groq React D3.js Docker

πŸ›‘οΈ Multimodal Deepfake Detection via Synchronization Fusion

Undergraduate Research @ MSRIT (Adv: Dr. S. Sethu Selvi)

Portfolio Link

Identifies synthetic facial video manipulations by capturing temporal-phase mismatches between acoustic formants and visual phoneme movements.

πŸ”¬ Innovations & Benchmarks:

  • Cross-Modal Attention: Synchronization-Aware Feature Fusion (SAFF) combining EfficientNet-B0 visual streams with Mel-spectrogram CNN audio encoders via CM-GAN.
  • State-of-the-Art Results: Achieved 94.57% accuracy and 0.979 AUC on the LAV-DF benchmark dataset; demonstrated robust cross-lingual generalization on Hindi datasets (HAV-DF: 80.12% accuracy).

Stack: PyTorch EfficientNet-B0 Mel-Spectrograms CM-GAN OpenCV Librosa

πŸ” Pramanik AI: SOC 2 Multi-Modal RAG Engine

Autonomous Security & Compliance Infrastructure

Source

Makes compliance self-serve for SaaS startups through a multi-modal RAG architecture operating across 7 specialized compliance modes (SOC 2, ISO 27001, HIPAA).

πŸ”¬ Innovations & Benchmarks:

  • Zero-Lockin Runtime Failover: Dynamic runtime provider failover across DeepSeek β†’ AWS Bedrock β†’ Groq.
  • CES Risk Vectorization: Custom risk vectorization algorithm with hybrid dense-sparse semantic retrieval across Supabase vector tables.

Stack: Python FastAPI AWS Bedrock DeepSeek Supabase React Pydantic v2

πŸ“Š NexusAI: AI-Powered Application Tracking

Event-Driven Job Search Intelligence Platform

Live Demo Source

Replaces static application tracking with automated Gmail OAuth synchronization, background queue processing, and structured LLM extraction.

πŸ”¬ Innovations & Benchmarks:

  • Non-Blocking Ingestion: Background BullMQ + Redis job consumers decoupling OAuth token refresh and rate-limited LLM inference.
  • Hybrid Extraction: Dual-engine RegEx pre-filtering + Gemini 1.5 JSON-schema validation for deterministic telemetry and analytics.

Stack: React 18 TypeScript Node.js BullMQ Redis MongoDB Gemini 1.5


πŸš€ More Production Builds & Systems

Project Description Highlights Stack
🩺 HealTrip Medical Tourism Microservices Platform 3 independent ML engines, clinical PDF NLP for specialty detection, automated visa letter generation FastAPI Node.js scikit-learn MongoDB Stripe Docker
πŸ› BetterFeedback Automated Feedback Classification Triages user feedback into Bugs / Features / Pain Points with sentiment scoring; 22 tests in 0.14s Python Flask Pydantic v2 React SQLite
βš™οΈ text_to_verilog Natural Language to Hardware Description Translates natural language module specifications into syntactically valid Verilog RTL modules Python LLM Verilog Hardware Synthesis
πŸ” research-qa Semantic Document Intelligence Vectorized retrieval-augmented generation engine with dense embeddings for academic paper interrogation Python RAG NLP FAISS
🌐 Academic Portfolio Minimalist Academic Research Portal Full-width single-column responsive dashboard; custom hash router; zero external framework bloat Vanilla JavaScript CSS3 HTML5 GitHub Pages

πŸ› οΈ Technical Arsenal

Languages & Core Computation

Python C++ TypeScript JavaScript SQL

Machine Learning & Signal Processing

PyTorch Hugging Face scikit-learn FAISS Weights & Biases

Distributed Systems, Backends & Data

FastAPI Node.js Redis Docker PostgreSQL MongoDB

Frontend & Data Visualization

React Next.js D3.js Tailwind Three.js


πŸ“Š GitHub & Algorithmic Activity




🧠 Research & Engineering Principles

Principle How I Practice It
πŸ“ Statistical Rigor First Validate agent convergence and model lifts with empirical regret bounds and Wilcoxon signed-rank testing ($p < 0.001$).
⚑ Decoupled Cascades Route inputs through fast heuristic filters before committing to computationally heavy Transformer sequence inference.
πŸ”„ Async & Non-Blocking Background task queues (BullMQ/Redis), streaming WebSockets, and event-driven decoupling across all services.
πŸ›‘οΈ Zero-Leakage Preprocessing Strict deduplication and GroupKFold partitions to eliminate cross-fold contamination on real-world distributions.
πŸ”’ Type Safety Everywhere End-to-end TypeScript on client interfaces + strict Pydantic v2 schemas on all backend microservices.

🎢 Beyond Engineering

When away from terminal screens and neural networks, I am a dedicated practitioner of classical Indian performing arts:

  • Hindustani Classical Vocal Performance: Trained across Khyal, Tarana, and semi-classical forms (Akhil Bharatiya Gandharva Mahavidyalaya Mandal certification); broadcasted on All India Radio (AIR) and Doordarshan.
  • Tabla Percussion: Rigorous study of rhythmic cycles (Teentaal, Ektaal), improvisational kaydas, and accompaniment dynamics.

πŸ“¬ Let's Connect


Academic Portfolio Β Β  LinkedIn Β Β  Email



Actively seeking & available for full-time Research Fellow / Intern appointments (6–10 months) and Tier-1 AI Systems roles.


Pinned Loading

  1. NexusAI NexusAI Public

    TypeScript 1

  2. HealTrip HealTrip Public

    JavaScript 1

  3. pramanik pramanik Public

    JavaScript 1

  4. research-qa research-qa Public

    Python 1