cv

This page contains my detailed CV, highlighting my background. You can download a PDF version using the button above.

Contact

Name Arnav Rastogi
Email arnavrastogi.543@gmail.com
Website https://arnav.rastogi.net.in/
LinkedIn arnav-rastogi-6b868a22a
GitHub ARC345

Research Interests

  • Interpretability, Recursive Models, Novel Model Architectures, BitNets.

Education

  • 2024 - present
    MicroMasters
    MITx
    Statistics and Data Science
    • Coursework: Probabilistic Modeling, Machine Learning, Time Series Analysis.
  • 2025 - 2029
    BS
    Indian Institute of Technology, Jodhpur
    Applied AI and Data Science
    • CGPA: 8.63/10
    • Coursework: Deep Learning, Linear Algebra, Statistical Inference.

Projects

  • 2026
    Learning & Unlearning Noise in State-Space Models ([code](https://github.com/ARC345/learn-unlearn-mamba), [write-up](https://arnav.rastogi.net.in/blog/2026/mamba-noise-learning-unlearning/))
    • Extended a QUEST Lab (IISc) study on noise learning/unlearning in transformer SLMs to Mamba-1.4B, running the full three-phase protocol (clean finetune → noisy finetune → clean retrain) on SQuAD across 4 noise types.
    • Found Mamba absorbs token-level noise far more than transformers (71% accuracy drop under word-flip vs. 27–38% for Olmo/Qwen/Gemma/Phi-2), while counterfactual noise raised its accuracy, a behavior no transformer showed.
    • Identified confounds limiting the comparison (base vs. instruction-tuned models, fuzzy-match vs. LLM-judge scoring, extra epochs inflating Phase 3 recovery to 124–133%) and documented them alongside results.
  • 2023
    ARC-Engine
    • Developed a high-performance OpenGL rendering engine in C++ with a custom Entity Component System (ECS) for cache-friendly memory layout.
    • Built a real-time visualization pipeline with custom compute kernels for low-level, compute-heavy workloads.
  • 2025 - present
    rCiv (Knowledge Commons)
    • Built a FastAPI service where LLM agents submit source-referenced claims that resolve into a temporal knowledge graph (Graphiti + Neo4j), with LLM-based deduplication and conflict detection over hybrid semantic/BM25/graph retrieval.
  • 2025 - present
    LLM Lab
    • Implemented a character-level GPT from scratch in PyTorch (multi-head attention, residual stream, LayerNorm) as a testbed for architecture experiments.

Technical Writing

  • Multi-Hop Reasoning in Transformers (link) Jan 2026
  • ReLU vs GELU (link) Jan 2026

Experience

  • 2026 - 2026

    Pune, India

    AI Engineer Intern
    Hapticware Intelligence
    • Built an agent orchestration framework with LangGraph coordinating 30+ agents across decomposable tasks.
    • Engineered RAG over a 100+ table production database, cutting chatbot query latency by 30-40% and auto-generating result graphs.
  • 2025 - 2025

    India

    Research Intern
    Abhay Travels
    • Built a CatBoost demand-forecasting pipeline with 15+ engineered seasonal features that beat ARIMA baselines on RMSE over 5 years of booking data.

Skills

  • ML & Research PyTorch, Transformers, State-Space Models (Mamba), LLM finetuning & evaluation, CatBoost
  • Mathematics Linear Algebra, Probability & Statistics, Optimization
  • Engineering Python, C++, SQL, OpenGL, FastAPI, Git

Open Source

  • 2026
    OmniRoute ([PRs](https://github.com/diegosouzapw/OmniRoute/pulls?q=is%3Apr+author%3AARC345+is%3Amerged))
    • 3 merged PRs to a 74K-star open-source LLM gateway, including Zed-hosted provider auth for model discovery, provider wire-format fixes, and repairing two live-server E2E CI suites.