BE Computer Engineering @ PESMCOE, SPPU · CGPA 9.34/10
Building AI systems across LLM infrastructure, machine learning, and autonomous robotics.
Structural code retrieval for LLM coding agents — zero-LLM index, served over MCP.
SkeletonGraph is a structure-aware retrieval layer for coding agents. It builds a zero-LLM repository index from source code structure and exposes targeted retrieval through the Model Context Protocol (MCP), so an agent lands on the exact function to edit instead of reading whole files to find it.
| Index | tree-sitter symbol graph + call graph + PageRank centrality (10 languages, no LLM calls) |
| Retrieval | BM25 + code embeddings + structural rerank, fused via reciprocal rank fusion |
| Interface | MCP server + portable sg CLI |
| Evaluation | SWE-bench Verified & SWE-rebench, 100-task Docker-verified deployment study |
Highlights
- Zero-LLM indexing — repository structure built with no API calls, instant to rebuild after every edit
- Matched a frontier coding agent's native solve rate while cutting p95 (tail) cost by 42%
- Hybrid lexical, semantic, and structural retrieval — highest recall of 8 benchmarked pipelines
- Cross-file call graph + PageRank centrality built through tree-sitter parsing
- Published to PyPI and the official MCP Registry — works with Claude Code, Cursor, Copilot, Windsurf, and any MCP client
Constraint-preserving hierarchical memory for long-horizon LLM conversations.
HierMem is an OS-inspired hierarchical memory architecture designed to preserve constraints and relevant context across long conversations. Inspired by virtual-memory paging, it combines a lightweight curator, a priority-aware constraint store, and a multi-level memory hierarchy to manage context under a bounded token budget.
| Curator | Reads a compact L0 index to identify relevant memory |
| Retrieval | Keyword · Semantic · Hierarchical · Hybrid |
| Assembly | Attention-aware bounded context construction |
| Memory | Constraint extraction and hierarchical L0 → L3 archival |
| Evaluation | Result |
|---|---|
| Memory compression | 4.7× vs. raw context |
| Constraint survival | 93.3% over long sessions |
| LLM-as-Judge | 8.4–8.7 / 10 |
| Evaluated against | RAG · RAG + Summary · MemGPT-style · Raw context |
|
Step-level diagnosis and targeted fine-tuning for LLM reasoning failures.
|
Deepfake detection designed for robustness across changing data distributions.
|
|
Weather-aware ship-route optimisation using adaptive path planning.
|
End-to-end machine-learning pipeline for real-time fraud scoring.
|
|
Multimodal AI platform for food analysis and nutrition-oriented insights.
|
Multi-agent governance simulation with adaptive decision policies.
|
Software & AI Lead
Developing autonomous navigation and perception systems for PESMCOE's ABU Robocon 2026 robots.
- Path planning under partial observability and dynamic obstacles
- Computer vision and real-time perception pipelines
- ROS2-based software architecture
- Simulation and validation using NVIDIA Isaac Sim and Gazebo
- Integration with embedded controllers and physical robot hardware
ROS2 NVIDIA Isaac Sim Gazebo Python C++ Computer Vision
| Languages | Python · C++ · TypeScript · SQL · MATLAB |
| Machine Learning | PyTorch · Transformers · QLoRA/PEFT · XGBoost · SigLIP · sentence-transformers |
| LLM Systems | Ollama · llama.cpp · GGUF · LiteLLM · RAG · MCP · tree-sitter |
| Robotics | ROS2 · NVIDIA Isaac Sim · Gazebo · Path Planning · Reinforcement Learning |
| Backend & Data | FastAPI · Flask · PostgreSQL · MongoDB · Firebase |
| Cloud & Tools | Google Cloud · AWS · AMD Developer Cloud · Docker · Git · pytest |
"The gap between research and production is where I work."