Data Scientist and AI Systems Engineer with an M.Sc. in Data Science from the University of Kalyani, specializing in production-ready AI that bridges advanced academic research and enterprise engineering.

I drive technical innovation across dual capacities: as a Data Scientist at Insightrix Consultancy, and as a core engineer within the RKMV R&D Cell in industry collaboration with FaceOff Technology. Through these institutional roles and specialized consulting engagements, I architect end-to-end intelligence systems—spanning speech forensics, multimodal deepfake detection, agentic RAG, and real-time telephony analysis.

Published researcher at ACL 2025 and CLEF 2025. Competed globally in multilingual NLP shared tasks.

"Whether leading R&D initiatives or consulting for commercial clients, I build systems that listen, reason, and deliver actionable intelligence—at scale and in real time."

• Offline-first target voice extraction from degraded wiretap-quality recordings using speaker embeddings with adaptive quality gating
• Dual-pass heuristic scanning (600s sparse → 1.5s dense) for multi-hour inference with dynamic peak/sustained threshold enforcement
• Automated speaker segmentation with diarization-based identity canonicalization and cosine drift prevention
• Court-admissible evidence packaging — isolated WAV clips + JSON metadata + timestamped PDF forensic reports
• Browser-based forensic audio editor with multi-region waveform editing, effects chain, and undo/redo state management

PyTorch SpeechBrain ECAPA-TDNN Silero VAD pyannote Librosa Django Flask Celery Redis

PyTorch WavLM Wav2Vec2 MediaPipe ViT OpenCV FastAPI Twilio Django Channels WebSockets Docker

• Multi-task audio deepfake detector with learnable 12-layer aggregation, Transformer encoder, and jointly predicted real/fake + codec + attack-family classification heads
• Unified training across 4 benchmark datasets with mislabeled sample correction (+43% real data recovery) and production-grade augmentation (Opus/AMR codec, WebRTC effects, MUSAN noise)
• 6-class speech emotion recognition (9,587 utterances, F1=0.71) with layer-wise attention and telephony-aware augmentation
• Real-time telephony fraud screening — live call analysis with deepfake risk scoring, emotion inference, and multilingual forensic summaries
• Multimodal meeting security — concurrent audio + video analysis over WebSocket streams with temporal smoothing and risk-level alerts

• Session-aware AI assistant for audio/video case investigation with analysis-conditioned RAG prompting and persistent video-linked chat history
• Real-time voice-to-voice AI assessment platform with async orchestration running STT, speculative RAG pre-fetching, and continuous forensics as concurrent tasks
• Tri-modal reasoning engine fusing audio forensics (Wav2Vec2), text forensics (LLM linguistic analysis), and visual forensics (micro-expression/gaze/posture) into structured agentic decisions
• Bidirectional PCM audio streaming via Gemini Live API with auto-reconnection, fire-and-forget filler phrases, and sub-200ms perceptual latency

FastAPI WebSockets Gemini 2.5 Gemini Live API Milvus Wav2Vec2 Whisper Django asyncio

Qwen-32B-AWQ DeepSeek-R1-7B llama.cpp Milvus BGE-M3 BGE-Reranker HNSW Django

• Dual-LLM agentic architecture — Qwen-32B-AWQ (GPU, Flash Attention 2) as reasoning LLM + DeepSeek-R1-7B (CPU via llama.cpp) as critic model with singleton pattern
• 3-stage hybrid retrieval: bi-encoder semantic search (HNSW on Milvus) → cross-encoder re-ranking → smart context stitching at sentence boundaries
• ReAct agentic loop with custom tool registry, multi-angle evidence injection, thought-based loop detection, and fallback summaries
• Query Cognition layer — LLM-powered intent classification, domain-jargon translation, and adaptive retrieval strategy injection
• Deployed over streaming NDJSON endpoints with real-time reasoning trace visualization and session-aware memory

Audio feature engineering (MFCCs, spectral contrast, chroma) + ML/DL for mood prediction. ~83% accuracy with Random Forest.

Python Librosa scikit-learn

Large-scale customer classification using MLP in PySpark with automated feature transformations and distributed preprocessing.

PySpark MLP Big Data

End-to-end fraud detection workflow with feature engineering, Random Forest achieving 94% accuracy, and precision/recall tuning.

Python scikit-learn Pandas

LLM-powered Reddit bot with OpenAI GPT for intelligent comment generation and real-time automated response loops.

Python OpenAI PRAW