Independent AI Researcher · LLM Memory Architectures · Open Source Contributor
Building novel memory architectures for LLMs — independently, without lab resources.
- DSPM (Dynamic Semantic Patch Memory) — 82.4% token reduction, 94.2% compression-retention ratio across 7 technical dialogue domains, statistically validated (Wilcoxon signed-rank, p = 0.0431) · Zenodo preprint · Accepted into the Sarvam AI Startup Program
- SSPM (Streaming Sparse Prompt Modeling) — real-time evaluation and constraint-preservation methodology validated on the LoCoMo benchmark · Zenodo preprint
- Vyom OS (ALISA — AI-Integrated OS Architecture) — systems-level architecture spanning bootloader, paging, and priority scheduling · Presented, under review, IEEE-affiliated ICSEAI 2026
- Yale
pie(pie-project/pie) — 4 merged PRs: #382 (stderr forward-port), #383 (sharded GGUF loading), #391 (E2E test coverage for raw-completion), #407 (async system test coverage) · Issue #418 — isolated a Vulkan-backend memory-corruption defect via clean baseline comparison, documented workaround - Point Cloud Library (PCL) — Merged PR #6398 · Fixed non-deterministic azimuth orientation in
ShapeContext3DEstimation - Pyserini / Anserini (Castorini Research Group, University of Waterloo) — Merged PR #2483 and PR #3227, reviewed and merged by Prof. Jimmy Lin; reproduced and validated baseline metrics (BM25, uniCOIL, SPLADE, Contriever) in a controlled Ubuntu 24.04 / Java 21 environment
- UW VibeServe — PR #57, referred by Prof. Luis Ceze, reviewed by Prof. Baris Kasikci
Python · C/C++ · PyTorch · OS Kernels & Memory Management · LLM Inference Pipelines · REST APIs · PostgreSQL · Supabase · Groq API · Gemini API
- 300+ LeetCode problems solved (Arrays, Linked Lists, STL, DP, Graphs, Trees)
- 331 combined Zenodo views, 186 downloads across two published preprints
- NVIDIA GTC 2025 attendee · Solo participant, Meta PyTorch Hackathon × Scaler
Portfolio · LinkedIn · LeetCode · ORCID · dhruvdubey1311@gmail.com