Union Command Agent: A 6.44M-Parameter English-Hinglish Parser for On-Device Command Execution

This research technical note describes Union Command Agent, a 6,441,472-parameter encoder-decoder Transformer trained from scratch to translate English and Hinglish device-control requests into a constrained action language. The system supports a catalog of 371 actions and uses a NumPy CPU inference runtime with grammar-constrained decoding. It is a task-specific command parser; execution policies and user confirmation remain separate from model prediction. An audit of the archived predictions gives 10,445/10,618 exact matches (98.37%) on the main test set and 6,448/6,611 (97.53%) on its reported strict subset. Fresh runtime checks reproduce 369/371 curated matches and 461/461 Wi-Fi matches. Existing device benchmark records report median parse latencies of 18.4 ms on an Intel Core i7-1360P laptop and 30.2 ms on a Raspberry Pi 5; these saved measurements were not rerun for this publication. The release preserves the eight-page manuscript, original model/checkpoint archive, current associated training and evaluation archive, source and reproduction scripts, model and dataset cards, numerical evidence, and file hashes. The training snapshot contains 1,230,162 rows. The author confirms that the supplied A/B collections and augmentation rows are original or synthetic work. No historical run-time hash proves that this current archive is byte-identical to the data used in the original training run; that provenance limitation is explicitly documented. Limitations include development-time evaluation exposure, a small number of exact train/evaluation overlaps, broader template overlap, no matched baseline or multi-seed study, and incomplete rejection of unsupported requests: only 316/400 main-test unknown examples were correctly rejected. Parser accuracy does not establish execution safety or general autonomous-agent capability. This technical note has not been peer reviewed. AI assistance in auditing and manuscript preparation is disclosed in the paper. Model: sraivante/tiny-superfast-agentic-MLM-6.5m-v1Versioned dataset: tiny-superfast-agentic-MLM-6.5m-v1-data, v1.0.0 Licensing by artifact: manuscript and research narrative: Copyright (c) 2026 Lalit Belwal, CC BY 4.0. Original model, code, dataset contributions and numerical evidence: Copyright (c) 2026 sraivante, Apache License 2.0. See LICENSES.txt for scope and retained third-party terms.