B.Tech (Hons.) CSE with Major in AI & ML, Manipal University Jaipur
AI Researcher @ IAIRO | Ex-Apprentice @ IISERK | Ex-Researcher @ Pragya Lab, BITS Goa | Ex-ML Researcher @ IIT-H Vigil Labs | NeurIPS, CoRL & WACV Workshop Papers | Springer Nature Published Author | 3× Dean's List
Who I am:
- 🔬 AI Researcher @ IAIRO (Neurosymbolic AI, Knowledge Graphs, Embodied AI)
- 🧪 Ex-Apprentice @ IISERK (SAEs and Mechanistic Interpretability), Ex-Researcher @ Pragya Lab, BITS Goa (NLP & LLM Evaluation), and Ex-ML Researcher @ IIT-H Vigil Labs (Federated Learning)
- 📄 Author — 3 workshop papers at NeurIPS 2026, CoRL 2026 & WACV 2027, plus journals with Springer Nature: SN Computer Science [Q1 - Top 25% in field] and Discover Computing [Q2]
- 🌱 Selected for the inaugural, fully funded cohort of the Agyeya AI Safety & Interpretability Workshop, IISc Bangalore
- 🏆 3× Dean's List — Excellence in Academics & Off-campus Achievements
- 🎓 Graduated B.Tech (Hons.) CSE with Major in AI & ML @ Manipal University Jaipur
What I build:
- 🧠 Building at the intersection of ML, NLP, Computer Vision, and Generative AI
- 🔍 Working with Neurosymbolic AI, Knowledge Graphs, SAEs & Mechanistic Interpretability, VLMs & LVLMs, Embodied AI, Federated Learning, LLM fine-tuning
- 📈 Exploring Model Inference, model optimization, and LLM deployment at scale
Let's connect:
- 🤝 Open to collaborations on AI research, impactful projects, and roles in Applied ML.
|
Proposes answerability, a fixed-reader metric that separates the evidence a scene-graph memory retains from what a QA model guesses from priors. Four targeted repairs (higher-resolution sensing, closed-vocabulary prompting, per-type calibration, door-segmented rooms) double object recall (24.1% to 49.2%), recover the visited-room count in 199/200 episodes, and raise answerability from 27.2% to 37.4% and retrained QA from 71.4% to 75.9%, while showing that higher object F1 can lower answerability. Keywords: Embodied AI, Scene Graphs, Episodic Memory, VLMs, Embodied QA, Evaluation |
|
Introduces a graph-swap test that tells apart a VLM that receives its memory from one that relies on it. Fine-tuning Qwen2.5-VL-7B (LoRA + R-GCN), text memory adds +30.1 pts on route questions while standard fusion stays at no-memory level; a per-node token route gives a swap-verified +8.3 pt graph-only gain and exposes a shortcut: next to text, the model ignores its graph. Keywords: Scene Graphs, GNNs, VLMs, Shortcut Learning |
Builds a multi-room EQA benchmark (5,700 episodes, 1,563 ProcTHOR houses, trajectory-linked symbolic QA) with a 7-rung evidence ladder. A closed-loop Qwen3-VL agent reaches 0/200 targets yet matches a no-image baseline (45.5%), versus 57.4% under oracle replay; the audit surfaces 3 shortcut classes. Keywords: Embodied QA, Benchmarks, Neurosymbolic AI, Evaluation |
|
Integrates DenseNet and EfficientNetB3 for hierarchical feature extraction with XGBoost as the final classifier. Class-Selective Image Preprocessing (CSIP) emphasises tumour regions, Grad-CAM localises discriminative features, and a HIPAA and GDPR-compliant AWS pipeline enables real-time clinical use, with federated learning provisions for multi-institutional privacy. Keywords: Histopathology, Ensemble Learning, Grad-CAM, AWS |
A lightweight boundary-aware strategy that emphasises contour pixels during local training with zero added communication overhead. Across 6 architectures on ISIC 2018 with 100 federated clients under a Dirichlet non-IID split, it cuts HD95 boundary error by 14% and reaches the highest Dice (87.63%) and IoU (80.06%). Keywords: Federated Learning, Segmentation, Non-IID, HD95 |
|
ML system for barbell exercise classification and repetition counting from MetaMotion sensor data, reaching over 90% accuracy through feature engineering and outlier detection for precise human activity recognition. |
|
Production-ready local RAG Text-to-SQL chatbot that converts natural language into safe PostgreSQL queries with 84% SQL accuracy (80% end-to-end) and ~2.5s latency. Features schema-aware retrieval (ChromaDB), code-specialized LLMs via Ollama (Qwen2.5-Coder / CodeLlama), SQL validation & security hardening, and 64 automated tests with RAGAS evaluation. Stack: Python, PostgreSQL, Ollama, Chainlit, ChromaDB, nomic-embed, pytest |
|
High-performance FastAPI microservice achieving 90% reduction in prediction latency (50ms → 5.4ms). Improved False Positive precision from 6% to 78% using class-weighted XGBoost evaluated via MLflow. Full production observability via Prometheus + Grafana, Dockerized for AWS deployment. Stack: Python, FastAPI, XGBoost, MLflow, Docker, AWS, Prometheus, Grafana |
RAG pipeline with sub-second query latency and 95% relevance accuracy across 12,000+ anime entries using LangChain, Groq LLM, and ChromaDB. Automated ETL with HuggingFace Sentence Transformers. Deployed on GCP with Kubernetes, maintaining 99.5% uptime with Grafana monitoring. Stack: Python, LangChain, Groq LLM, ChromaDB, Streamlit, Docker, Kubernetes, GCP |
🔬 Research Experience
🏛️ AI Researcher — Indian AI Research Organization (IAIRO) (Mar 2026 – Present)
- Built a neurosymbolic knowledge-graph memory for long-horizon embodied question answering (ICLR 2027, under review), spanning a 5,700-episode multi-room ProcTHOR benchmark, VLM scene graphs merged into episodic knowledge graphs, and LoRA fine-tuning of VLMs on H200 GPUs
- Extended this work into 3 workshop papers at NeurIPS, CoRL and WACV within 3 months of joining the project, under Dr. Amit Sheth (first author on 1, co-first on 2)
🏛️ AI Research Associate — IISER KOLKATA (Apprenticeship) (Mar 2026 – Oct 2026)
- Researched Sparse Autoencoders (SAEs) and Mechanistic Interpretability
- Investigated internal feature representations to understand and decompose LLM behavior
🏛️ NLP Research Intern — Pragya Lab, BITS Goa (Jun 2026 – Oct 2026)
- Researched NLP and LLM evaluation
- Worked on evaluation methodologies for large language model behavior and outputs
🏛️ ML Research & Development Intern — VIGIL Labs, IIT Hyderabad (Apr 2025 – Jul 2025)
- Engineered a decentralized Federated Learning model for medical image classification & segmentation
- Surpassed baseline test accuracy by 20% on complex non-IID real-world medical data
- Reduced global communication round time by 45% and utilized 55% fewer resources than baselines
- Orchestrated secure ML workflows to address data heterogeneity and strict distributed data privacy requirements
🏆 Honors & Awards
- 🧠 Selected for the inaugural, fully funded cohort of the Agyeya AI Safety & Interpretability Workshop, IISc Bangalore (2026)
- 🤖 3 workshop papers at NeurIPS 2026 (NEmo), CoRL 2026 (SISREN) & WACV 2027 (HOUR)
- 📄 3× Springer Nature Published Author — SN Computer Science [Q1 - Top 25% in field], Discover Computing [Q2] & ICDEC 2024
- 🎓 Dean's List for Excellence in Academics (Highest GPA) — Manipal University Jaipur
- 🏅 2× Dean's List for Excellence in Off-campus Achievements — Manipal University Jaipur
- 🤝 Strong academic foundation in CS & AI combined with hands-on industry and research impact
☁️ MLOps & Deployment Expertise
End-to-end experience taking AI projects from development to large-scale production:
- MLOps: MLflow, DVC, DAGsHub, Airflow, Astro Airflow, TaskFlow, BentoML, Grafana
- LLMOps: LangSmith, LangServe, LangGraph, LangChain
- Cloud: AWS (S3, EC2, IAM, RDS), GCP, Azure
- Containers: Docker, Kubernetes
| Category | Tools / Frameworks |
|---|---|
| AI & LLMs | GPT, BERT, Titan, Hugging Face Transformers, LLM Fine-tuning, RAG Pipelines |
| Deployment | Flask, Streamlit, FastAPI, BentoML |
| Visualization | Tableau, PowerBI, Plotly, Seaborn, Grafana |
| Dev Tools | Git, GitHub, Docker, Postman, MLflow, DVC, DAGsHub, Selenium |
| Collaboration | Jira, SCRUM, Agile |
| Domains | ML, DL, NLP, CV, Generative AI, LLMs, Federated Learning, MLOps, LLMOps, Healthcare AI |
If you find my projects helpful or interesting:
- ⭐ Star my repositories
- 🔔 Follow me on GitHub for updates
- 🤝 Collaborate on AI/ML research and projects
- 💬 Share with others who might benefit
💭 "Render thy labor as sacred offering unto the Most High; and lo, thy toil shall cease to be burden, becoming instead eternal joy"
⭐ If you find my work interesting, consider starring my repositories!
🤝 Open to collaborations, research opportunities, and roles in ML/AI/Healthcare
Made with ❤️ by Divyansh Pandey
💼 Seeking opportunities in world of applied ML/NLP/CV research | 🌐 Let's build something amazing together!

