Looking for a job

Fedra Yuny

@fedrayuny

Message

I build high-performance LLM inference, AI agents, RAG systems, and open-source models.

Brazil

Message

I've built open-source LLM infrastructure, agent runtimes, RAG applications, models, and datasets. My work spans GPU/CPU inference, distributed systems, persistent agent state, and reproducible benchmarks.

With MegaGemm, I built continuous batching, KV caching, and 160+ Triton/CUDA kernel execution paths for Qwen 3.5 and Gemma 4. Its FP16 benchmark reached 93.1% of vLLM 0.27.1's geometric mean throughput on Qwen 2.5 3B/L4, while MicroGemm reached 1.286× llama.cpp's geometric mean throughput in a paired CPU benchmark.

I created JadeAgent and J.A.D.E./JadePort to support tool calling, graph workflows, recovery, multi-agent execution, model routing, and persistent memory. I also built Lancer for cited web research and AetherMap for hybrid retrieval, where hybrid retrieval with reranking achieved MRR 0.99 and Hit@3 100% on SQuAD-PT.

As an independent researcher, I trained GTLM, an approximately 2B-parameter MoE, and released nine Portuguese conversational models through JadeLLM. I also created a Brazilian Portuguese corpus of 8.4 million documents and approximately 6.29 billion tokens, alongside 993,000 synthetic QA pairs for LLM and RAG work.

Experience

Work history, roles, and key accomplishments

Current

Independent Researcher

Current

GTLM

Jan 2025 - Present (1 year 8 months)

Designed and trained an approximately 2B-parameter MoE with 350M active parameters per token from scratch on a single A100. Achieved a 56.2% average across six zero-shot benchmarks, exceeding reported averages for Pythia-410M and GPT-2 Medium.

MOE Pretraining PyTorch A100 Zero Shot Benchmarks Apache 2.0

Creator and Lead Architect

MegaGemm

Built an LLM inference platform with continuous batching, KV cache, and 160+ Triton/CUDA kernel execution paths. Developed distributed inference for heterogeneous hardware and a CPU runtime with INT8/INT4 and AVX2/NEON.

LLM Inference Triton CUDA C C++Continuous Batching KV Caching Distributed Inference Heterogeneous Hardware INT8 INT4 AVX2 NEON

Machine Learning Engineer

AetherMap

Developed hybrid RAG with FAISS, BM25, and reranking, plus a custom fine-tuned embedding model. Built an evaluation suite with over 1,000 documented query executions and five ablation modes.

Hybrid RAG Faiss BM25 Reranking Embedding models Evaluation LLM As Judge Profiling

Creator and Architect

JadeAgent

Built an agent runtime with tool calling, graph workflows, and execution policies, plus asynchronous mesh execution with capability-based routing. Designed an execution state format with persistence and recovery.

Agent Runtime Platforms Tool calling Graph Workflows Asynchronous Processing Capability Based Routing State Management Recovery SQLite File Based Data Persistence

Creator and Developer

Lancer

Built research workflows with query decomposition, concurrent searches, and cited reports streamed via SSE. Implemented reranking combining semantic relevance, domain authority, and recency, plus autonomous browsing with Browser Use and E2B.

AI Powered Web Research Query Decomposition Concurrent Searches SSE Reranking Bi Encoder CrossEncoder Browser Use E2B CDP Rate limiting

Architect and Full-Stack Developer

J.A.D.E. / JadePort

Developed a FastAPI platform with routing across multiple LLM providers, model fallback, and session management. Implemented an agent with ReAct, multistep planning, and semantic code retrieval, plus RAG over PDFs and video transcripts.

fastAPI LLM routing Fallbacks Session Management React Planning Semantic Code RAG ChromaDB GitHub integration E2B

Creator and LLM Engineer

JadeLLM

Released nine models ranging from 0.6B to 72B parameters using SFT, LoRA/DoRA, distillation, and adapter merging. The ecosystem reached 35,000 downloads, including 26,600 for community quantizations.

SFT LoRAs DORA Distillation Adapter Merging Model Release Quantization

Creator

Corpus PT-BR & RAG QA

Built a corpus of 8.4 million documents and approximately 6.29 billion tokens, with curation, deduplication, and Parquet publication. Generated a derived dataset of 993,000 synthetic QA pairs using Mistral.

Data Engineering Curation Deduplication Parquet Synthetic Data Mistral