AI/ML Engineer · LLM Infrastructure · Robotics · Production ML

Rising 4th-year BE Computer Engineering @ PESMCOE, SPPU · CGPA 9.34/10

Building AI systems that work outside the lab — from research prototype to deployed package.

---

🧠 Flagship Research — HierMem

LLMs forget. HierMem fixes that.

HierMem is an OS-inspired hierarchical paged memory architecture for long-horizon LLM conversations. Inspired by virtual memory paging — the same design that lets your computer run programs larger than physical RAM — it uses a stateless curator agent, a priority-tagged constraint store, and a 4-level memory hierarchy (L0→L3) to prevent context degradation, constraint forgetting, and hallucination.

┌─────────────────────────────────────────────────────────────┐
│  Curator (~1000 tokens, constant)  →  reads L0 index only   │
│  Retrieval  →  KEYWORD / SEMANTIC / HIERARCHICAL / HYBRID   │
│  Assembler  →  4-zone attention-optimised context           │
│  Main LLM   →  generates from bounded ~6000-token budget    │
│  Post-Proc  →  extracts constraints, updates L0→L3 archive  │
└─────────────────────────────────────────────────────────────┘

Metric	Result
Memory compression ratio	4.7× vs raw context
Constraint survival rate	93.3% over long sessions
LLM-as-Judge score	8.4–8.7 / 10 (vs 5.4–7.6 baselines)
Outperforms	RAG · RAG+Summary · MemGPT-style · Raw LLM

---

🔬 Projects

LLM Reasoning Pipeline Step-level diagnosis and targeted fine-tuning for LLM reasoning failures — diagnoses where models fail, not just if they fail. Fine-tuned Qwen 2.5 3B with QLoRA 4-bit 63% reduction in step-level failure rate Ground-truth backtracking + error propagation analysis Runs fully local via llama.cpp — zero API cost PyTorch QLoRA PEFT GGUF Ollama	AI-Generated Video Detection Deepfake detection pipeline designed to stay robust across domains and changing distributions. 92% AUC-ROC across DFDC, CelebDF, WildDeepfake SigLIP transformer embeddings + blockwise frame representation Continual learning via replay buffers + knowledge distillation Prevents catastrophic forgetting across dataset shifts PyTorch SigLIP Transformers Continual Learning
Fraud Detection System Production-grade fraud detection — not a demo, a deployed pipeline. 96.2% accuracy, 0.92 AUC-ROC Processes 50K+ transactions/day XGBoost + neural networks with advanced feature engineering Real-time FastAPI classification with geolocation tracking XGBoost FastAPI Neural Networks sklearn	Self-Evolving Multi-Agent Governance Decentralised agents that negotiate and evolve decision policies autonomously. PPO + quadratic voting with reputation weighting Agents negotiate rule proposals in a simulated digital economy PostgreSQL logging of governance events and adaptive reward strategies Delivered as a functional 24-hour hackathon proof-of-concept Ray RLlib PettingZoo PPO PostgreSQL
AquaIntel ⭐ 7 Ship routing system with live weather integration. Enhanced A* algorithm with adaptive grid weighting Integrates wind, depth risk, and fuel efficiency in real-time Reduced route update time from 3.6h → 36 minutes (6×) Full-stack with route visualisation via Leaflet.js TypeScript A* Algorithm OpenWeatherMap Leaflet.js	NutriSense AI health platform built around generative AI orchestration. Gemini 2.0 Flash for multimodal food recognition NLP-driven dietary Q&A + longitudinal health trend analysis Serverless GCP backend with Firestore real-time persistence WCAG 2.1 accessible UI with Firebase Auth + server-side token validation Gemini API Google Cloud Firebase Vanilla CSS

---

🤖 Robotics — ABU Robocon 2026

Team Vulcans · Software & AI Lead

Building the autonomous navigation and AI stack for PESMCOE's entry in ABU Robocon 2026.

• Modified A* planning for partial observability and dynamic obstacles
• PPO-based control with reward shaping for rule-compliant navigation
• ROS2 architecture + Gazebo simulation environments
• Validated in NVIDIA Isaac Sim before real-world hardware integration

ROS2 NVIDIA Isaac Sim Gazebo PPO Python MATLAB

---

⚙️ Tech Stack

LANGUAGES   = ["Python", "TypeScript", "C++", "SQL", "MATLAB"]

ML_AI       = ["PyTorch", "Hugging Face TRL", "QLoRA/PEFT", "LiteLLM",
               "LangChain", "XGBoost", "sentence-transformers", "SigLIP"]

LLM_INFRA   = ["Ollama", "llama.cpp", "GGUF", "ChromaDB",
               "RAG pipelines", "LLM-as-Judge evaluation"]

ROBOTICS    = ["ROS2", "NVIDIA Isaac Sim", "Gazebo", "A*", "PPO", "Ray RLlib"]

BACKEND     = ["FastAPI", "Flask", "PostgreSQL", "MongoDB", "Firebase"]

CLOUD       = ["Google Cloud Platform", "AWS"]

TOOLS       = ["Git", "Docker", "pytest", "Streamlit", "Weights & Biases"]

---

📊 GitHub Stats

---

📈 Contribution Activity

---

🎯 Currently

• 🔬 Extending HierMem — Needle-in-a-Haystack and multi-step reasoning benchmarks
• 🤖 Autonomous navigation stack for ABU Robocon 2026
• 📖 Rising 4th year BE CS · Open to research internships & AI engineering roles (2025–26)
• 📬 Reach me: [email protected] · linkedin.com/in/yash-doke

---

"The gap between research and production is where I work."