Al-ameen Olajide

software engineer based in lagos, nigeria

this is where i write up the work in depth - training and quantizing open models, agent memory, inference runtimes, and the evaluation that decides whether any of it actually held. long posts, real numbers, reproducible where i can make them.

self-taught since 2020. currently studying computer science at lagos state university.

now building bad theory labs - an independent ai research lab in lagos shipping open models, agent infrastructure, native runtimes, and reproducible research. the principle: build ambitious systems, measure them without mercy, ship what survives.

our latest is btl-4, a 35b open-weight agentic reasoning model trained on an execution-gated corpus - trajectories were kept only where the code actually ran and passed its tests. 78.4% on swe-bench verified, 73.5% on bfcl v4, 262k native context, apache 2.0.

btl-4 compact is the same model in a single 9.96gb file at 2.30 bits per weight, retaining 94.1% of full-precision behaviour. it loads in llama.cpp, ollama, and lm studio - a frontier-class coding model on hardware you already own. btl runtime is our openai-compatible inference api with multi-provider routing.

retaindb is the persistent memory layer for agents - memory that survives across sessions, grounded retrieval, sub-40ms lookups. state of the art on longmemeval preference recall.

previously a full-stack systems engineer at costcrunch (nyc, remote) where i architected enterprise-grade multi-tenant infrastructure with zero-trust security architecture, designed high-performance semantic search systems processing massive datasets with sub-millisecond latency, and engineered fault-tolerant event-driven systems with kafka.

specialize in distributed systems architecture, real-time data processing, blockchain security protocols, and production-scale machine learning infrastructure.

Contact