AHMAD GARBA ADAMU/AKA AGA/@RUFATRONICS

NIGERIA → WORLDWIDESYS.ONLINE

AI Engineering & Research

I BUILD
INTELLIGENT
SYSTEMS.

AI systems, edge-first software, and experiments built for real-world constraints. Focusing on low-resource machine learning, cross-platform agents, and African-focused technology.

VIEW WORK EXPLORE AI LAB

[EDGE AI][LOW-RESOURCE ML][NLP & ROUTING][GEOSPATIAL INTEL]

CURRENTLY BUILDING

LIVE

VELKRO / EDGE AI

Autonomous agent architecture & 35M-parameter edge classifier pipeline.

STATUSACTIVE

FOCUSLOW-RESOURCE ML

LOCATIONNG / GLOBAL

SYSTEMONLINE

LATENCY: ~2msMEM: 2GB ADAPTIVE

COMPUTATIONAL INTERFACESparse node geometry & local edge execution.

01 // INDEXED PORTFOLIO

SELECTED WORK

Engineering archive featuring geospatial intelligence, local browser AI, self-hostable agent daemons, 35M intent routers, and edge agricultural vision.

01GEOSPATIAL AI / DEFENSE / SIGINT[2025]

NORTHWATCH TACTICAL / NW0

High-fidelity geospatial intelligence platform for predictive threat analysis across Nigeria.

Northwatch is a unified geospatial intelligence HUD built for predictive situational awareness, tactical reasoning, and threat monitoring across Nigeria's 37 sector boundaries.

// KEY ARCHITECTURE & FEATURES

›National AI Grid with 5,400+ point density mesh
›Mapping across 37 sectors & Critical Flashpoints
›Pulsing HUD markers & threat classification
›Tactical reasoning / SITREPs & SIGINT feed
›Tactical Mode & High-res Satellite Mode rendering

React 18ViteTypeScriptLeafletSupabase RealtimeTailwind CSSFramer Motion

GITHUB REPO LIVE DEMO

SYS_RENDER // NORTHWATCH-TACTICALSTATUS: OK

TACTICAL HUD // GEOSPATIAL MESHA unified predictive operating picture designed to move from reactive reporting toward proactive intervention.

GRID MESH5,400+ points

SECTORS37 covered

PROTOCOLSITREP-4.1

ENCRYPTIONAES-256-GCM

02EDGE AI / WEB AI / PRIVACY[2025]

DEYSAFE

Lightweight privacy-first web AI suite engineered for 2GB RAM edge devices.

DeySafe proves useful AI should not require an expensive flagship device or permanent 5G connectivity. It executes computer vision and language inference locally inside the browser memory footprint.

// KEY ARCHITECTURE & FEATURES

›On-device TFLite vision inference & model lifecycle manager
›Browser-based WebLLM chat running via Web Workers
›Offline model caching via IndexedDB & Service Worker
›Low-memory load/unload/cleanup routines targeting ~2GB RAM
›Zero server roundtrips — full client-side execution

TypeScriptReactViteTensorFlow.jsTFLite WebMLC WebLLMIndexedDBService Worker

GITHUB REPO LIVE DEMO

SYS_RENDER // DEYSAFESTATUS: OK

DEYSAFE // ON-DEVICE INFERENCE ENGINEUseful AI should not require an expensive device or permanent cloud connection.

MEMORY TARGET~2GB RAM

LATENCYClient Native

STORAGEIndexedDB Cache

03AI AGENTS / GO / SYSTEMS[2025]

VELKROGO

Self-hostable cross-platform AI agent written in Go with explicit security policy controls.

VelkroGo combines two operative personas: Coder (development services & repo modification) and Operator (files, shell, browser, network). Every action flows through explicit, revocable approval gates.

// KEY ARCHITECTURE & FEATURES

›T0–T4 risk classification & policy gates
›Scoped session grants for high-risk capabilities
›Daemon architecture with Bubble Tea TUI and Web GUI frontends
›Provider-agnostic routing (Ollama, Anthropic, OpenAI)
›Self-modification workflow: Diff review → Self-test → Snapshot → Rollback

GoBubble TeaAnthropic APIOpenAI-compatible APIsOllamaWeb GUI

GITHUB REPO

SYS_RENDER // VELKROGOSTATUS: OK

VELKROGO // CODER & OPERATOR DAEMONHigh-performance low-resource architecture design

LANGUAGEGo (Golang)

RISK TIERST0 to T4

STATUSPhase 1 Implemented

04TRANSFORMERS / NLP ROUTING / ML RESEARCH[2025]

UEG / UNIVERSAL EDGE GATEWAY

35M-parameter bidirectional transformer for sub-3ms request routing before frontier LLMs.

UEG classifies incoming user intent, required resource class, and language across 22 classes before invoking expensive LLMs. It routes requests through 5 cost and compute tiers.

// KEY ARCHITECTURE & FEATURES

›22 intent classes & 8 supported languages
›5 routing tiers: DROP/BLOCK, STATIC, DEVICE API, CACHE, FRONTIER
›Trained on 176,000+ curated synthetic & real prompt examples
›Published open HuggingFace model and dataset
›Low latency lightweight classifier footprint

PyTorchHuggingFace TransformersONNXFastAPIDataset Curation

GITHUB REPO LIVE DEMO HUGGINGFACE

SYS_RENDER // UEGSTATUS: OK

UEG // 35M PARAMETER INTENT ROUTERHigh-performance low-resource architecture design

ACCURACY~90%+ (Reported)

AVG LATENCY~2ms (VS 220ms Llama)

PARAMETERS35M Transformer

COVERAGE100%

INTERACTIVE CLASSIFIER DEMO

UEG-35M Intent & Routing Engine

MODEL: UEG-35M|LATENCY: ~2ms

Test how the 35M-parameter bidirectional transformer evaluates intent, selects local vs cloud execution tiers, and calculates routing decisions prior to LLM invocation.

TRY PRESETS:

// CLASSIFICATION OUTPUTPROCESSED IN 2.1ms

PREDICTED INTENTAGRI_DIAGNOSTICS_QUERY

RESOURCE CLASSLOW_RAM_VISION_MODEL

LANGUAGE / DIALECTHAUSA / ENGLISH

CONFIDENCE SCORE98.4%

TIER 3 — DEVICE APIExecutes on local device NDK pipeline without cloud LLM invocation.

OPTIMAL ROUTE

05COMPUTER VISION / ANDROID NDK / EDGE AGRI[2024]

FARMBOT

Android native crop diagnostic system designed for low-tier hardware in African agriculture.

FarmBot brings computer vision diagnosis directly to rural farmers. Running natively via C++ NDK on Android 5.0+ devices, it functions completely without internet connection.

// KEY ARCHITECTURE & FEATURES

›MinSdk 21 (Android 5.0+) target for older phone compatibility
›Native C++ inference pipeline via CMake and NDK bindings
›Offline crop disease detection & treatment recommendations
›Companion HuggingFace models: crop-assistant, crop_doctor, health_scan
›Interactive Web Gradio demo on HuggingFace Spaces

KotlinAndroid NDKC++CMakeTFLiteGradioGitHub Actions

GITHUB REPO HUGGINGFACE

SYS_RENDER // FARMBOTSTATUS: OK

FARMBOT // NATIVE C++ DIAGNOSTIC PIPELINEBring AI-powered crop diagnostics to farmers even when they have old phones or unreliable internet access.

MIN SDK21 (Android 5.0)

INFERENCENative C++ NDK

OFFLINE100% Autonomous

02 // EXPERIMENTAL DIVISION

AI LAB & RESEARCH

MODELS / DATASETS / BENCHMARKS / EDGE SYSTEMS

Active research on low-resource LLM routing and edge quantization.

// CATEGORY:

EXPERIMENT / ENTRY

TYPE

RESEARCH SCOPE

TECHNICAL SPECS

STATUS

VELKRO-LM SPECIFICATIONS

ID: velkrolm-arch

MODEL

Compact sparse-attention transformer research exploring token efficiency and computational graphs on arm64 SoC clusters.

#Transformer#Sparse Attention#Edge

Custom Kernel / Arm64 NEON

RESEARCH

UEG 176K INTENT DATASET

ID: ueg-dataset-176k

DATASET

176,000+ multi-turn and single-turn intent prompts across 22 operational classes in 8 major African & international languages.

#NLP#Multilingual#Dataset

176,400 Rows / JSONL / Open-Source

PUBLISHED

CROP DISEASE TLM DATASET

ID: crop-disease-tlm

DATASET

Thermal and multi-spectral leaf anomaly dataset gathered specifically for low-resource West African agricultural contexts.

#Computer Vision#Agriculture

Multi-class Disease Taxonomy

PUBLISHED

UEG-35M TRANSFORMER

ID: ueg-35m-classifier

MODEL

Sub-3ms classifier model optimized with ONNX runtime for rapid edge-to-cloud request dispatching.

#Routing#ONNX#Edge

35M Params / FP16 Quantized

PUBLISHED

ARM64 INFERENCE BENCHMARKS

ID: edge-latency-bench

BENCHMARK

In-depth latency vs RAM tradeoff study on MediaTek Helio G-series and Snapdragon 6xx chipsets for TFLite Web and WebLLM.

#Benchmarking#Android NDK#TFLite

2GB - 4GB RAM Hardware Profiles

ACTIVE

LOCAL & EDGE INFERENCE THESIS

All models and datasets produced under the Rufatronics initiative prioritize transparent benchmarks, open-access weights, and functional low-RAM deployment.

HUGGINGFACE ORGANISATION

03 // TECHNICAL CREDIBILITY

CAPABILITY MATRIX

Engineering domains and core stack utilized across edge systems, research experiments, and production deployments.

AI / ML & TRANSFORMERS

■PyTorch
■HuggingFace
■Transformers
■ONNX Runtime
■Intent Routing
■Model Quantization (FP16/INT8)
■Dataset Curation

EDGE INFERENCE & LOW-RESOURCE

■TFLite Web / Native
■TensorFlow.js
■MLC WebLLM
■Android NDK (C++/CMake)
■Local Memory Lifecycle
■IndexedDB / Service Worker

SYSTEMS & AGENT ARCHITECTURE

■Go (Golang)
■Bubble Tea TUI
■Security Policy Gates (T0–T4)
■Provider-Agnostic LLM Routing
■Autonomous Workflows
■Diff Review & Rollbacks

GEOSPATIAL & WEB APPLICATIONS

■TypeScript
■React 18
■Vite
■Tailwind CSS
■Leaflet / React-Leaflet
■Supabase Realtime Postgres
■Framer Motion

04 // HUMAN ENGINEERING NARRATIVE

ABOUT AHMAD

KANO / ABUJA, NIGERIA → WORLDWIDE

AHMAD GARBA ADAMUAGA / @RUFATRONICS

SYS.VER: 2026

LAT: 12.0022° N, LON: 8.5920° EEDGE AI / RESEARCH

Building practical technology under real-world computational constraints.

I am Ahmad Garba Adamu (known online as @rufatronics or Aga), an AI engineer and low-resource systems researcher based in Nigeria.

My work centers on the intersection of machine learning, edge computing, and low-resource systems. I build tools that operate under strict bandwidth, memory, and hardware constraints—ensuring that cutting-edge artificial intelligence serves people in environments where expensive GPUs and permanent 5G connectivity are not guaranteed.

From on-device crop diagnostics running natively in C++ on older Android phones (FarmBot) to privacy-first browser AI suites (DeySafe) and sub-3ms intent classification models (UEG), my engineering philosophy prioritizes utility, local execution, and open research.

EDGE-FIRST ML

Prioritizing local on-device inference, quantization, and offline resilience.

AFRICAN CONTEXTS

Engineering specifically for low-RAM mobile hardware and regional needs.

05 // DIRECT COMMUNICATIONS

HAVE SOMETHING
WORTH BUILDING?

Available for technical consultations, edge AI engineering projects, research collaborations, and open-source contributions. Direct contact methods below:

// PRIMARY EMAIL DISPATCH

// PUBLIC PROFILES & REPOSITORIES

github.com/rufatronics

huggingface.co/rufatronics

linkedin.com/in/ahmad-garba-adamu

NO MIDDLEMEN:Direct communication preferred. Responses within 24 hours.