Debjyoti GhoshAI/ML Developer & Student at UEM Kolkata
AI/ML Developer · Student at UEM Kolkata · Python Backend Engineer
I am an undergraduate student pursuing B.Tech in CSIT at the University of Engineering & Management (UEM), Kolkata, and an AI/ML developer & Python backend engineer. Formerly Backend Engineer at ThreatPurge. Published researcher at IEEE INDICON 2025 specializing in agentic AI trust, multimodal deepfake detection, voice biometrics anti-spoofing, and scalable microservices.
Skills & Technologies
Tools, frameworks, and programming languages I use to build research models and production backend infrastructure.
My Research Work
Cutting-edge research in AI/ML with focus on deepfake detection, forensics, and biometric applications
Multimodal Deepfake Detection via Audio-Visual Fusion
95% (0.97 AUC) Accuracy
Architected a scalable multimodal deepfake detection pipeline utilizing a Transformer-based cross-attention mechanism to fuse spatial, temporal, and acoustic embeddings, achieving 95% accuracy and 0.97 AUC.
Key Highlights
- Integrated ResNet-50 visual encoder, BiLSTM temporal encoder, and 1D-CNN audio encoder processing Mel-spectrograms.
- Implemented custom synchronization loss function to enforce phoneme-viseme alignment verification.
- Evaluated on FakeAVCeleb benchmark dataset demonstrating superior cross-modal robustness.
Technologies
Sonar: Real-Time On-Device Speech Intelligence & Diarization
2,433.5 ms (93.3% Speaker Accuracy)
Engineered an edge speech intelligence system running entirely on a Raspberry Pi 5. Combines USB audio ingestion, Silero VAD v5, dual-stream ECAPA2 diarization with AOSC adaptation, and faster-whisper INT8 ASR.
Key Highlights
- Fully autonomous on-device deployment on Raspberry Pi 5 with zero cloud dependencies or external network calls.
- Verified live hardware trial: processed 49.2 seconds of conversational speech across 15 utterances with 2,433.5ms total latency.
- AOSC exponential moving average profile adaptation dynamically tracks acoustic drift as speakers move relative to the microphone.
- Dual-stream architecture synchronizing CTranslate2 INT8 ASR with SpeechBrain ECAPA2 512-D speaker embeddings via SCDiar aligner.
Technologies
DAVAI: Autonomous Agentic Runtime with 4-Tier Memory & Dynamic Tool RAG
0ms Cache Hits (42.1% TTFT Reduction)
Architected an autonomous, stateful desktop AI agent runtime decoupling LLM reasoning from execution via an asynchronous EventBus, dynamic multi-model routing (Groq, DeepSeek, Ollama), and zero-delay WebSocket streaming.
Key Highlights
- Formulated a 4-tier hybrid memory engine fusing ephemeral RAM working memory (45m sliding TTL), ChromaDB vector embeddings, 2-hop Knowledge Graph triples, and SQLite (WAL mode) ground truth.
- Engineered Dynamic Tool RAG 2.0 with a balanced-brace nested parser, dynamically pruning 19+ tool schemas to top candidate functions using hybrid vector and lexical retrieval.
- Implemented 10-layer deterministic prompt sequencing and an ephemeral LRU cache (<50MB RAM), stabilizing prefix bit-patterns to maximize cloud KV prompt cache reuse.
- Constructed multimodal deep research and vision pipelines integrating headless Playwright Chromium web crawling with targeted Tesseract OCR heuristics.
Technologies
Elaina: Enterprise Multi-Agent Legal Intelligence Platform
48 Verified E2E API Ops Accuracy
Architected a multi-tenant Agentic AI platform for legal matter management, featuring dual-path orchestration (Fast Path conversational agent vs Slow Path background intelligence queue), Bayesian Neural Network risk calibration, and human-in-the-loop review queues.
Key Highlights
- Engineered dual-path orchestration: sub-650ms Fast Path real-time SSE chat and asynchronous background intelligence queue.
- Implemented Bayesian Neural Network (BNN) uncertainty calibration providing confidence bounds for legal outcome assessments.
- Integrated human-in-the-loop attorney audit queue ensuring zero unverified citations reach client-facing documents.
- Verified with comprehensive 48-endpoint automated E2E API integration suite.
Technologies
Context-Aware Hate Speech Detection via Post-Comment Joint Modeling
95% (0.9911 AUC) Accuracy
Created a context-aware moderation framework replacing isolated text analysis with post-comment joint modeling, achieving 95% accuracy and ROC-AUC of 0.9911.
Key Highlights
- Structured four-class reaction taxonomy mapping discourse to explicit moderation actions.
- Fused parallel features including SBERT semantic similarity, emotion distributions, and DistilBERT probabilities.
- Validated framework robustness on curated social media interaction dataset.
Technologies
