# Debjyoti Ghosh — AI/ML Developer & Student at UEM Kolkata

> Debjyoti Ghosh is an AI/ML developer, Python backend engineer, and CSIT student at the University of Engineering & Management (UEM), Kolkata. Formerly Backend Engineer at ThreatPurge. Published researcher at 22nd IEEE INDICON 2025 (IIIT Bangalore) specializing in agentic AI trust, multimodal deepfake detection, voice biometrics anti-spoofing, and scalable microservices.

- **Full Name**: Debjyoti Ghosh
- **Location**: Kolkata, India
- **Email**: debjyoti1ghosh@gmail.com
- **Website**: https://www.debjyoti-ghosh.in/
- **GitHub**: https://github.com/debjyoti71
- **Google Scholar**: https://scholar.google.com/citations?user=hg2SqvMAAAAJ&hl=en
- **LinkedIn**: https://www.linkedin.com/in/debjyoti-ghosh-508bba281/
- **Education**: B.Tech in Computer Science & Information Technology (2023–2027), University of Engineering & Management (UEM), Kolkata (CGPA: 8.7/10)

---

## Quick Navigation & Core Canonical Routes

- [Homepage](https://www.debjyoti-ghosh.in/): Primary portfolio overview with research pillars, interactive DAVAI agent terminal, and verified credentials.
- [Projects Archive](https://www.debjyoti-ghosh.in/projects): Complete archive of published peer-reviewed research papers and production open-source systems.
- [Experience & Industry Roles](https://www.debjyoti-ghosh.in/experience): Professional track record at ThreatPurge Private Limited (FastAPI Backend / Anti-Phishing) and Elaina Legal AI.
- [Achievements & Honors](https://www.debjyoti-ghosh.in/achievements): IEEE publications, national hackathon podiums, robotics championships, and Amazon ML Summer School.

---

## Featured Research Projects & Case Studies

- [Multimodal Deepfake Detection via Audio-Visual Fusion](https://www.debjyoti-ghosh.in/projects/multimodal-deepfake-detection): IEEE INDICON 2025 publication. Cross-attention Transformer framework fusing ResNet-50 visual, BiLSTM temporal, and 1D-CNN Mel-spectrogram acoustic embeddings. Evaluated on FakeAVCeleb, achieving 95% accuracy and 0.970 ROC-AUC with phoneme-viseme synchronization loss. DOI: 10.1109/INDICON68490.2025.11392906.
- [Context-Aware Hate Speech Detection via Post-Comment Joint Modeling](https://www.debjyoti-ghosh.in/projects/context-aware-hate-speech-detection): Accepted research paper. Solves isolated comment moderation hallucinations by joint-modeling root post and reply threads using SBERT and DistilBERT. Achieves 95% accuracy and 0.9911 ROC-AUC with a 4-action moderation taxonomy (Approve, Flag, Soft-Warn, Auto-Quarantine).
- [Edge-Based Voice Biometric Verification & Cloud Anti-Spoofing](https://www.debjyoti-ghosh.in/projects/voice-biometrics-anti-spoofing): Audio forensics architecture combining edge-based 13-dim MFCC feature extraction with cloud FastAPI anti-spoofing verification and spectral jitter defense. 97.5% authentication precision.
- [DAVAI Autonomous Agentic Runtime](https://www.debjyoti-ghosh.in/projects/davai-agentic-runtime): Stateful desktop AI agent runtime with asynchronous EventBus, dynamic multi-model routing (Groq, DeepSeek, Ollama), 4-tier hybrid memory (ephemeral RAM, ChromaDB vector embeddings, 2-hop Knowledge Graph, SQLite WAL), and Dynamic Tool RAG 2.0 reducing TTFT by 70%.
- [AntiGravity (Dev Timekeeper)](https://www.debjyoti-ghosh.in/projects/antigravity-dev-timekeeper): Local-first VS Code and VSCodium productivity extension published on Open VSX Registry with 1,300+ downloads. Zero telemetry leakage.

---

## Publications & Primary PDF Preprints

- [IEEE INDICON 2025 Paper (PDF)](https://www.debjyoti-ghosh.in/papers/Deepfake_Detection_Using_Transfer_Learning.pdf): "Multimodal Deepfake Detection via Audio-Visual Fusion", 22nd IEEE INDICON 2025, IIIT Bangalore.
- [Hate Speech Moderation Paper (PDF)](https://www.debjyoti-ghosh.in/papers/Hate_Speech_Detection_Reaction_Classification.pdf): "Context-Aware Hate Speech Detection via Post-Comment Joint Modeling".
- [Voice Biometric Forensics Report (PDF)](https://www.debjyoti-ghosh.in/papers/Voice_Security_Project_Report.pdf): "Voice Biometric Authentication & Cloud Anti-Spoofing Architecture".

---

## Complete Context & LLM Grounding

For the full textual corpus including complete abstracts, step-by-step methodologies, evaluation benchmarks, BibTeX citations, and technical skill taxonomy, refer to:
- [Full LLM Knowledge Base (llms-full.txt)](https://www.debjyoti-ghosh.in/llms-full.txt)