About Me
Who I am and what drives me
Auishik Pyne
AI agent engineer. Making Meta AI Glasses smarter at Reality Labs. Building a self-hosted autonomous agent for a secure digital life. Pursuing MSCS @ Georgia Tech. Obsessed with AI that's actually useful in the real world.
I work at the intersection of machine learning, system design, and practical application — from evaluating multimodal LLM outputs for Meta AI to architecting a modular "second brain" that manages tasks, calendar, and knowledge autonomously. My focus is on building AI that isn't just impressive in a demo, but genuinely improves how we live and work.
Work Experience
Where I've worked and what I've built
Data Quality Analyst @ Meta
- Graded multimodal & voice Meta AI LLM responses against strict guidelines, maintaining >95% quality accuracy across thousands of data points.
- Evaluated complex reasoning tasks for Meta AI Glasses, surfacing model quality gaps to improve on-device AI performance.
- Audited Bengali ↔ English conversational translations for Meta AI training pipelines, sustaining >96% contextual accuracy and linguistic fidelity.
- Led peer review sessions to support DQA certification prep, strengthening team-wide annotation consistency and performance.
QA Specialist @ Prosrvc LLC
- Processed and inspected 100+ devices daily with accuracy and attention to detail.
- Maintained clear documentation and flagged errors in real time, ensuring smooth workflow.
- Supported team efficiency by organizing workstations and coordinating quality reports.
Sales Associate @ The TJX Companies, Inc.
- Helped customers, answered questions, and handled cash register transactions accurately.
- Kept the store clean, organized displays, and promoted credit cards, consistently passing monthly goals.
Software Engineer @ REVE Systems
- Deployed a gender and age predictive model from speech, achieving 93% and 74% accuracy respectively.
- Optimized a custom-trained VITS text-to-speech model for ONNX on Bengali datasets, doubling inference speed for cross-platform deployment.
- Developed a Your TTS-based multi-speaker model enabling diverse voice synthesis with a single model.
- Enhanced ASR by training NVIDIA NeMo QuartzNet on Bengali datasets with beam search.
- Built a Speaker Diarization system using Nvidia TitaNet-L, identifying speakers in multi-party conversations with >80% accuracy.
- Created a 3× faster FastSpeech 2 pipeline integrating TorchServe, WebSocket, and Redis for efficient batch processing.
- Developed a tri-layered Biometric Authentication System with speaker, facial, and fingerprint verification, backed by ROC & AUC analysis.
Education
Academic background
Master of Science in Computer Science
Specializing in Machine Learning. Fall 2025 intake. Focus areas include deep learning, reinforcement learning, and autonomous agent systems.
Bachelor of Science in Electrical & Computer Engineering
Built a strong foundation in theoretical knowledge and practical skills across computer programming, data structures, algorithms, DSP, neural networks, and computer vision.
Projects
What I'm building in my free time
Self-Hosted AI Agent — Second Brain
Building a modular, self-hosted AI agent that autonomously manages tasks, calendar, email, and knowledge. Designed a modular LLM adapter layer enabling hot-swapping between local models (DeepSeek via Ollama) and cloud providers (Claude, GPT-4) without changing downstream logic. Integrated ChromaDB vector memory for persistent semantic recall across sessions.
The Medical Bill Crusher
Architected an AI-powered medical bill auditing pipeline that automates discrepancy detection between medical bills and EOBs (Explanation of Benefits) to generate regulatory appeal scripts, helping patients catch billing errors.
Interactive PDF Chat Application
Developed a source-aware RAG (Retrieval-Augmented Generation) pipeline implementing advanced chunking and semantic search techniques to minimize hallucinations and significantly improve context retrieval accuracy for document Q&A.
AI Data Analytics Agent
Built a conversational agent to automate data analysis, allowing non-technical users to query datasets using natural language and receive insights, visualizations, and actionable summaries without writing code.
Bengali Text to Speech
Engineered a 3× faster FastSpeech 2 pipeline integrating TorchServe, WebSocket, and Redis for efficient batch processing of Bengali text-to-speech. Also optimized a custom-trained VITS model for ONNX deployment on Bengali datasets.
Gender & Age Prediction from Speech
Deployed a gender and age predictive model from speech achieving 93% and 74% accuracy respectively, using advanced digital signal processing and machine learning techniques on audio features.
Publications
Research contributions
Performance Comparison of Multiple Supervised Machine Learning Algorithms for COVID-19 Mortality Prediction
Explored how artificial intelligence aids in predicting COVID-19 patient outcomes. Leveraging machine learning models including Logistic Regression and K-Nearest Neighbor, this study achieved 95% accuracy and 89% F1 score. This vital research utilized the John Hopkins University dataset to enable impactful healthcare solutions for minimizing COVID-19 mortality.
Read Full Paper on IEEEExtracurricular
Beyond the code
Music & Singing
Passionate vocalist specializing in traditional Bengali music — including Rabindra Sangeet, Nazrul Sangeet, and Bengali patriotic songs. I play the harmonium and guitar and have performed at school, college, and university events.
Cricket & Badminton
Active sports enthusiast since childhood. Played on school cricket teams and local clubs, recognized for expertise in multiple positions. Also won awards at various levels in badminton competitions.
Community Service
Regular blood donor supporting local hospitals and clinics. Advocate for the importance of blood donation and have actively participated in blood donation camps and drives organized with health services.