AI Systems Engineer · Product Builder
AI Systems Engineer · Edge AI · Voice & Vision · Agentic Systems
I build multimodal AI systems that run in the real world-from edge devices and on-device inference to cloud agents, memory systems and full-stack products.
Currently developing a fire-safety AI system with trained vision models and edge-oriented inference, while leading technical direction for an AI assistant platform at Mako IT Lab.
60
public GitHub repos
20+
mobile, web and AI projects
2
verified pub.dev packages
727
pub.dev downloads, last 30 days
5+
years shipping products
Current R&D
Mako R&D combining trained vision models, deterministic safety checks, compact language-model reasoning and edge-oriented deployment.
FireSafeX Vision R&D
Training and evaluating YOLO-family vision models with dataset, inference and evaluation tooling for safety-device use cases.
AI Assistant Platform
Leading technical direction and integration for a separate assistant platform; internal product and architecture details remain private.
Train
Build and refine domain-specific vision and language-model datasets.
Evaluate
Test inference behavior and record reproducible quality and performance results.
Optimize
Quantize compact models and prepare artifacts for constrained-device testing.
Integrate
Connect vision, safety logic, reasoning and multilingual guidance into the wider system.
Evidence boundary
Training, export and integration artifacts are verified. Published throughput figures remain engineering targets until backed by reproducible device benchmarks.
Experience
Product engineering depth with increasing focus on AI systems, edge inference, multimodal interfaces and high-ownership product work.
2024 - Present
2023 - Present
Projects
Grouped by the kind of engineering signal companies can evaluate: AI systems first, then shipped AI products, then full-stack platforms.
Problem: Assistive navigation needs real-time obstacle understanding without a network connection.
Architecture: Offline camera-to-detector-to-spatial-reasoning-to-TTS pipeline with desktop and INT8 edge profiles.
Key achievement: Ongoing Mako R&D for safety devices, including YOLO-family model training; FPS figures are documented targets rather than published benchmarks.
Problem: Safety devices need modular AI pipelines that fit constrained compute, memory and power budgets.
Architecture: Split vision, safety validation and event-triggered small-model reasoning with LoRA, MLX, llama.cpp and quantized deployment targets.
Key achievement: Active Mako R&D across YOLO-family model training and constrained-device deployment planning.
Problem: LLM agents forget useful context or overfill prompts with noisy history.
Architecture: Semantic memory layer with compression, embeddings and retrieval for long-running agents.
Key achievement: Independently implemented the SimpleMem paper in TypeScript and published it as @sheriax/simplemem; benchmark figures belong to the cited paper.
Problem: Documents need searchable, grounded answers instead of generic chatbot responses.
Architecture: PDF ingestion, chunking, MiniLM embeddings, SQLite vector store and OpenRouter/DeepSeek answer generation.
Key achievement: Built a CLI RAG system with separate ingest and query commands.
Problem: Short-form video production has repetitive transcription, analysis and assembly steps.
Architecture: Whisper transcription, vision analysis and scripted media generation pipeline.
Key achievement: Automated core pieces of a TikTok/Reels/Shorts production workflow.
Problem: Financial recovery requires spending context, receipt understanding and practical guidance.
Architecture: Flutter app with Firebase, ML Kit OCR, rule-based tracking and OpenAI GPT for receipt scanning, advice, budgets, savings targets and alerts.
Key achievement: Built a production AI finance wellness app for spending tracking and financial recovery.
Problem: Users want a companion-like AI product that is simple enough for everyday use.
Architecture: Mobile AI companion with OpenAI/Google AI integrations and Firebase-backed user flows.
Key achievement: Built and shipped as a personal Android product, later folded into Sheriax and now discontinued.
Problem: Language-learning communities need safer content and lightweight AI assistance.
Architecture: Mobile learning app with moderation workflows and Firebase product infrastructure.
Key achievement: In-development Sheriax product with Youhana as sole developer.
Problem: Teams need local-first whiteboarding with reliable collaboration and sync.
Architecture: React canvas workspace with realtime collaboration, cloud sync and offline-first state.
Key achievement: Built a multi-board collaborative whiteboard platform.
Problem: A social-commerce product needed to evolve from its original cross-platform foundation into a richer video-first marketplace.
Architecture: React Native and Next.js in the Cubanin phase, followed by a full Flutter rewrite with Stripe Connect, S3 and Firebase as Echify.
Key achievement: Served as Tech Lead and Lead Mobile Developer, leading the full React Native-to-Flutter rewrite during the Echify rebrand.
Problem: Construction logistics needs offline support, GPS tracking and dispatch visibility.
Architecture: Offline-first mobile and web stack with maps, realtime tracking and route optimization.
Key achievement: Built logistics workflows around field reliability and realtime coordination.
Open Source
Small, practical libraries and tools around agent memory, Flutter payments, translation generation and cloud cleanup.
Lifelong memory for LLM agents with semantic compression and vector retrieval.
Flutter plugin for Stripe Connect embedded onboarding, payouts and payments UI.
Dart/Flutter translation key generation from JSON files with type-safe output.
Writing
Notes on the engineering problems behind real AI products: latency, memory, retrieval quality, voice UX, embedded constraints and agent reliability.
The real reason your AI prompts fail has nothing to do with prompting techniques — it's about how clearly you communicate as a person. Here's how learning to share goals instead of tasks changed everything for me, both with AI and with people.
AI agents forget everything between sessions. SimpleMem fixes that with semantic lossless compression — a three-stage pipeline that compresses, synthesizes, and retrieves long-term memories with 30× fewer tokens. Here's how @sheriax/simplemem brings this framework to TypeScript and Node.js.
Building a marketplace or platform app? flutter_stripe_connect is an open-source Flutter plugin that lets you embed all 13 Stripe Connect components directly into your Flutter app across iOS, Android, and Web — with just a few lines of code.
Toolkit
The technical areas, hardware and software I use when building AI systems and product infrastructure.