The most informative output of an autonomous pipeline is often what it refused to do. How the grad-agent makes identity errors structurally impossible, refuses when it cannot, and keeps a contestable record of every skipped lead.
Inside the Phrontlyne claims agent: how four fraud-detection layers carry Beta posteriors instead of fixed weights, how adjuster verdicts renegotiate the system's trust in its own components, and why 'is the learning loop actually closed?' is a first-class trustworthiness property.
A small honest writeup of two Twi English translation models published to Hugging Face, what worked, what did not, and what the numbers actually mean.
Design notes on Phrontlyne Claims Agent v2: a production insurance claims system with 45 MCP tools, two YOLO detectors joined by IoU geometry, a four-signal composite fraud pipeline, per-policy RAG, and a token efficiency layer.
A geospatial ML pipeline that combines Sentinel-2 imagery, NDVI, and a transformer object detector (RF-DETR) to track forest loss and illegal mining in Ghana.
A retrieval-augmented pipeline with citation validation and numeric grounding checks. Built as the second-round technical assessment for an Independent AI Engineer role at Deriv.com.
Introducing a bidirectional neural machine translation model for Twi and English, fine-tuned from MarianMT and trained on 10,000+ professionally translated sentence pairs.