Our Research. Our Mission.
We are an open-science artificial intelligence lab building optimized NLP translation engines, computer vision models, and low-latency API infrastructure.
The Democratization of AI
Traditional AI models are heavily skewed towards high-resource languages and commercial centers, leaving low-resource dialects underrepresented. At Pact AI (Mycl), we believe that linguistic inclusivity is critical to a connected future.
We compile regional datasets and fine-tune sequence-to-sequence transformer architectures, making high-quality neural machine translation viable for languages like Igbo, Yoruba, and other regional dialects.
Our operations are built on two pillars: Open Science (releasing models and weights publicly) and Privacy Compliance (never exploiting developer logs).
# Environment Specs
"framework": "PyTorch 2.3+CUDA 12.1",
"architecture": "T5-EncoderDecoder",
"quantization": "8-bit Integer (bitsandbytes)",
"distributed_nodes": "A100 SXM4 x 128",
# Optimization settings
"mixed_precision": "BF16 enabled",
"concurrency_limit": 4500,
"max_sequence_len": 512
Our Operating Principles
Releasing Weights
We distribute our optimized model checkpoints on Hugging Face so researchers globally can build on top of our datasets.
Inference Efficiency
We prioritize model quantization and weight distillation so complex models can compile efficiently on edge hardware.
Data Protection
We strictly separate user inference datasets from training nodes. Your API calls are private, audited, and purged.
High-Throughput Clusters
Our models compile on decentralized GPU configurations optimized for low latency and high availability. With automatic node load-balancing, Pact AI API clusters maintain 99.9% uptime.