Neural Quantization · Linear Transformers · Edge AI Systems
Nihal Gazi
AI Researcher & Systems Architect
I research extreme 1-bit neural quantization, linear-complexity vision transformers, and high-throughput zero-GPU edge inference systems.
One practice, three connected lanes
The work is organized by what it proves.
My research centers on 1-bit weight quantization, sparse and linear-complexity vision transformers, distributed vector caching, and CPU-only neural inference. I focus on removing hardware compute barriers so state-of-the-art models execute seamlessly on edge devices.
Resource-efficient neural systems and edge inference
Zero-GPU inference pipelines, first-principles architectures, and low-latency serving on commodity hardware.
Explore this lane 02 / Technical minorExtreme quantization and linear-time attention
1-bit weight binarization, sparse particle-tracking transformers, and distributed multi-tier caching architectures.
Explore this lane 03 / Public practiceOpen-source research and community communication
Technical sessions at GDG, public model releases on Hugging Face, and AI research at Pollinations.ai.
Explore this laneSelected technical work
Systems with a public record.
Realistic Text-to-Speech & FLUX Suite
Scalable community generative speech and vision pipelines on Hugging Face handling 840,000+ cumulative visits, 3,200+ combined likes, 42,000+ monthly active users, and recognized as Hugging Face Space of the Week.
Open repositoryFaceFlux: Zero-GPU Edge Facial Inference
Optimized CPU-only real-time facial inference pipeline achieving sub-second latency bounds; awarded National Top 10 standing and INR 1 Lakh research grant at SCNTSE 2025.
Open repositoryARGUS: AI Eye for CERN LHC Particle Tracking
Hierarchical Linear Vision Transformer replacing quadratic attention with strict O(N) localized projections for particle track reconstruction; published in EPJ Web of Conferences and awarded Conference Gold Medal.
Open repositoryOreoLook / search.elixpo
An open-source, OpenAI-compatible agentic web search and deep-research engine with Redis hot-caching and Qdrant semantic vector deduplication; yielded formal arXiv:2609.05463 preprint.
Open repositoryFast Color Transfer Algorithm (vecore)
Closed-form mathematical algorithm transferring chromatic profiles between high-resolution imagery with near-zero latency overhead; runs fully client-side in vanilla web technologies.
Open repositoryAzinet: Pure Java Neural Framework from Scratch
Zero-dependency neural network library engineered from first principles in pure Java (JDK 8) at age 17, implementing custom matrix math and backpropagation directly in bytecode arrays.
Open repositoryLinkedListTree: Custom Pointer-Based Tree Architecture
A custom hierarchical tree data structure implemented using linked list node references in Java, optimizing pointer-based traversals, dynamic node memory management, and tree balance.
Open repositoryEvidence, not a claim wall
Every role should connect to work someone can inspect.
Projects link to source. Research links to its record. Talks follow implemented lessons. Experience is described through responsibility rather than title alone.
How the work developedResearch and public explanation
Build first. Then publish the lesson.
ARGUS: The AI Eye for CERN — Solving Extreme Computational Physics
EPJ Web of Conferences, Volume 370, Article 01032 · Gold Medal (IEMPHYS-26)
Read the record Research / 2026A Three-Layer Caching Architecture for Low-Latency LLM Web Search on Commodity CPU Hardware
arXiv preprint · arXiv:2609.05463 [cs.DC], PollinationsAI Research
Read the record Talk / 2026Agentic AI Connect - 2026 Speaker
GDG JISU × ML Kolkata
View the session Talk / 2026Technical Hackathon Judge — Gemma For Bharat 2026
Google Developer Groups (GDG JISU)
View the sessionWork together