Hi, I'm Anuj Gupta 
Researcher & Developer · MCDS @ Carnegie Mellon University SCS (Aug 2025 – Dec 2026)
I build applied ML systems across LLMs, agentic pipelines, RAG and computer vision, from research prototypes to production on GCP and AWS. Right now I'm a graduate researcher at CMU's TEEL Lab and a technical contributor at Aspection Labs. Over summer 2026 I was a Senior Data Scientist Intern at Capital One.
- Language models for learning and education : research at TEEL Lab, CMU
- Video generation research and applied AI : technical contributor at Aspection Labs
| Role | Organization | When |
|---|---|---|
| Technical Contributor | Aspection Labs | Sep 2026 – Present |
| Graduate Researcher, TEEL Lab | Carnegie Mellon University SCS | Aug 2026 – Present |
| Senior Data Scientist Intern | Capital One | Jun 2026 – Aug 2026 |
| Graduate Researcher, TEEL Lab | Carnegie Mellon University SCS | Aug 2025 – May 2026 |
| Data Scientist | Equifax | Aug 2023 – Jul 2025 |
| Software Developer, Healthcare AI | Quantiphi | Aug 2022 – Jul 2023 |
| Founding Software Engineer | Expify | Jul 2021 – Jul 2022 |
| AI Research Intern, NLP & FinTech | Nippon Life India AMC | Mar 2021 – Jul 2021 |
Selected outcomes: cut AML case resolution time 98% (70 min → 2 min per case); 90% faster reporting across 300+ risk projects with LangGraph agents; credit-loss models at 87.6% accuracy informing over $45.5B in microfinance portfolios; 40% lower operational costs and 27% faster pipelines migrating legacy systems to GCP. Two Spotlight Awards at Equifax.
- Semantic Span Annotation (SSA): published at ACL 2026 (advisor: Dr. Jaromir Savelka). Unified span-annotation benchmark over 7 LLMs across 5 datasets, surfacing two distinct performance regimes: label definitions lifted F1 from 8.8% to 57.5% on ontology-heavy tasks but hurt pattern-based tasks like PII detection.
- RolloutKV (supervisor: Dr. Lei Li): profiled KV-cache rollout strategies for veRL GRPO post-training of Qwen2.5-Coder-3B on MBPP; cut training wall-clock from 12k to 9k seconds and raised rollout throughput ~1.56x.
- RedViz (advisor: Adam Perer): unified LLM red-teaming framework combining attention-map interpretability with realtime multilingual harm detection.
| Project | What It Does | Stack |
|---|---|---|
| EqRAG | LLM fine-tuning (LoRA, IA3) for Dow 30 stock prediction. 12% zero-shot to 51.7% accuracy with Qwen-2.5-Math-7B, training 0.19% of parameters | Python, PyTorch, HuggingFace, PEFT |
| RedViz | LLM red-teaming dashboard probing safety across 8 languages and 9 harm categories, with live jailbreak testing, token-level entropy, and attention maps on TinyLLaMA & LLaMA | Python, Streamlit, Altair |
| Skin-Burn-Detection | YOLOv5s classifying burn severity into 3 levels across 14,044 medically validated images, via transfer learning from COCO. Top 5 Finalist, TiE Global Hackathon (650+ teams) | Python, PyTorch, YOLOv5 |
| Anime-GAN | Three GAN architectures (DCGAN + Pix2Pix) generating anime faces from video frames. 28.5 dB PSNR, 0.91 SSIM | Python, Keras, PyTorch |
More Projects
| Project | What It Does |
|---|---|
| Credit-Card-Fraud-Detection | Fraud detection across 284,807 transactions at a 0.172% fraud rate; Random Forest reached 99.96% accuracy without resampling |
| Telecommunication-Churn-Prediction | Five classifiers compared on 7,043 telecom customers; neural network hit 0.83 AUC, with contract type the strongest churn signal |
- Google Cloud Certified : Associate Cloud Engineer
- AWS Certified : Solutions Architect Associate
