Professional Experience
Every role is a case study in driving business impact through AI.
Assistant Manager - AI Product Team
Anaptyss
Business Problem: The BFSI sector requires rigorous enterprise risk management (ERM) and control testing, which traditionally involves highly manual and slow audits.
Key Responsibilities: Leading a team of 8 engineers, architecting ANA (Adaptive Neuro Intelligent Assistant) utilizing a multi-agent system (14+ agents).
Impact & Metrics:
- Automated RCSA analysis and walkthrough scheduling via agentic workflows.
- Integrated ADK, tool calling, and TTS/STT for autonomous evidence validation.
- Spearheaded technical development of 3 end-to-end AI products.
Multi-Agent SystemsGoogle ADKTool CallingTTS/STTSystem Design
AI Engineer – Multi-Agent Systems
Anaptyss
Business Problem: Extracting complex financial covenants from thousands of legal documents with high accuracy and low latency.
Key Responsibilities: Architected an end-to-end multi-agent platform (CovenAce & Index-AI) for autonomous covenant extraction.
Impact & Metrics:
- Achieved 94.47% accuracy and 0.93 recall in entity classification processing 1000+ credit agreements.
- Reduced latency by 97% (5 hours down to 10 minutes) enabling real-time compliance monitoring.
- Improved extraction F1-score by 12% across diverse covenant types via aspect-based feedback mechanisms.
- Awarded Financial Express Award.
LLMsGoogle ADKPrompt OptimizationParallel Agent Architecture
ML Engineer – Computer Vision & Optimization
Ripik.AI
Business Problem: Industrial operations needed efficient edge-deployed computer vision and complex manufacturing scheduling optimization.
Key Responsibilities: Developing and optimizing CV models for edge deployment and building hybrid optimization engines.
Impact & Metrics:
- Compressed Segment Anything Model (SAM) reducing VRAM by 43% and latency by 32% while retaining 95% accuracy.
- Deployed production YOLOv8 fire detection system achieving 85% mAP@50.
- Increased client operational efficiency by 18% and net profit by 5% using Genetic Algorithms and Linear Programming.
- Implemented MLOps pipelines using Docker and AWS.
YOLOv8SAMONNXAWS EC2Genetic AlgorithmsMLOps
Research Intern – Multimodal Emotion AI
I'mBesideYou
Business Problem: Recognizing inner human emotions accurately from multimodal data sources in video streams.
Key Responsibilities: Developed multimodal late-fusion architectures combining text and audio, and facial landmark extraction pipelines.
Impact & Metrics:
- Achieved 93% accuracy on Japanese emotional speech dataset using BERT + Wav2Vec2.0.
- Reduced false positives by 23% using ROI-based temporal modeling.
- Awarded Letter of Recommendation.
BERTWav2Vec2.0Temporal ModelingMultimodal Late-Fusion