I build what others treat as a black box.

At Aden (YC W25), I engineer the MCP integration layer for Hivean autonomous agent framework designed to eliminate humans from the inference loop. My infrastructure enables AI agents to discover peer services, chain HuggingFace inference endpoints, generate embeddings, and recover from distributed failures autonomously.

I maintain 90%+ Pytest coverage on all agentic modulesbecause in distributed ML systems, a silent failure is not a bug. It's a catastrophe.

My research philosophy: you don't understand a model until you've derived its backward pass on paper. I implement foundational papersTransformers, Bahdanau Attention NMT, TextCNN, Word2Vecentirely in raw PyTorch. No wrapper libraries. No shortcuts.

Competitively, I hold LeetCode Top 16% Global (Rating 1655+, 530+ solved) and HackerRank 5-Star. I treat algorithms as the cognitive substrate for all engineering decisions.

Present

Open Source AI Engineer

Remote

Owner of the Model Context Protocol (MCP) integration layer for Hive, Aden's autonomous agent orchestration system. Engineering the full-stack infrastructure that enables AI agents to operate without human oversight at enterprise scale.

• Engineered MCP-based tool integrations enabling agents to autonomously execute ML models, generate real-time semantic embeddings, and orchestrate production HuggingFace Inference API endpoints.
• Architected authentication workflows, sliding-window rate limiting, and multi-layer fault-tolerance mechanisms for distributed agentic systems.
• Designed API coordination layers allowing autonomous agents to discover and call peer services, expanding Hive's capability for enterprise-scale multi-agent ML orchestration.
• Maintained strict 90%+ Pytest coverage with mocked responses and CI/CD integrity across all critical self-improving agentic AI infrastructure modules.

Full encoder-decoder Transformer for NMT. Scaled dot-product attention, parallel multi-head attention, sinusoidal positional encodings, dynamic batch masking, label smoothingall from scratch.

Original additive attention mechanism. Bidirectional GRU encoder + attention decoder. Hand-coded alignment model computing context vectors by scoring all source hidden states.

Parallel 1D conv filters over variable kernel sizes sliding across word embeddings, max-over-time pooling extracting critical n-gram features.

Word2Vec from zero. Skip-Gram + CBOW with negative sampling, full vocabulary pipeline, and noise-contrastive training loop. No Gensim.

NAXERION AI Multi-Modal Drug Recommendation System

Production-grade multi-modal medical AI fusing three foundation models via dynamic Softmax attention. RoBERTa for clinical text comprehension, Wav2Vec2 for acoustic emotion analysis, and CLIP for drug image verification. Achieves ~90% cross-modal diagnostic accuracy at sub-1.5s end-to-end latency. Also features a Gemma LLM fine-tuned via SFT + DPO and a DQN. Dockerized microservices with Kubernetes + NVIDIA GPU acceleration.

LEETCODE PRODUCTIVITY TRACKING SYSTEM

Event-driven Chrome Extension automating LeetCode-to-GitHub synchronization via GitHub OAuth 2.0 device flow. A background service worker intercepts LeetCode XHR submission responses and autonomously triggers REST API calls to push accepted solutions with structured problem metadata. Published as open-source developer tooling.

_{Automated Analytics Systems}

_{Full-Stack Dev · Feb 2026}

_{AI Innovation Track}