Machine Learning & Sequence Modeling (Hugging Face)

• parentheses-v0.9-130M-v3 (Project Tanzinite): 148.38M-parameter non-Transformer language model built from scratch on a single 8 GB consumer GPU (RTX 5050). It replaces quadratic attention with State Space Duality (SSD) linear recurrence: $O(T)$ chunked associative scan during training, and constant $O(1)$ memory state updates during inference with zero growing KV-cache. We added token-shift mixing ($\mu$-mix), $\beta_t$-gated write selectivity, and an FP32 recurrent highway to stop BF16 underflow across deep layers. Native script support covers Kannada, Tamil, Telugu, Malayalam, English, and low-resource Dravidian variants (Tulu, Koya, Kuvi, Kolami).
• parentheses-tokenizer-vocab: Custom 32,000 BPE vocabulary (v3_32k) with regex rules engineered specifically for Indic scripts. Instead of splitting multi-character aksharas into 3 to 5 byte tokens, it keeps consonants, viramas, and matras grouped together, cutting token inflation by 42% on Dravidian text. Includes <think>, </think>, and language routing tokens.
• dravidian-pretrain-sample: Decontaminated pre-training slice (Parquet) balanced across literary Dravidian texts, conversational dialog, and parallel translation bitexts.
• indic-reasoning-eval: Test suite containing verified prompts across arithmetic, multi-step deduction, and native-script Dravidian answers. Evaluates <think> reasoning tag closure and tracks bits-per-byte drift.

---

Systems, Kernels & Edge Tooling

• Parentheses (240M) (slab_v4): Our active 253.5M parameter recurrent model. Currently training on 655M cumulative tokens with textbook and code datasets, using custom Triton associative scan kernels and multi-head associative recall gates.
• Peridot: Air-gapped local AI operating kernel. It uses dynamic layer-by-layer VRAM arbitration and FAISS vector indices so a 27B model fits inside 8 GB VRAM without spilling unmanaged pages into system RAM.
• agent-wire: Model Context Protocol (MCP) server. Raw web pages waste up to 200k tokens per fetch. agent-wire parses the DOM locally and returns clean Markdown (70% to 85% token savings) with a local cookie vault so session tokens never touch external scraping APIs.
• ArgusCouncil: Multi-agent research coordinator with cryptographically signed provenance logs and isolated execution phases.
• Verity: Deterministic document-audit engine for internal accounts-payable controls. It executes 14 structural compliance checks on invoice data without trusting generative model outputs.
• Project Chakravyuh: Counter-fraud audio pipeline. It intercepts scam call traffic over WebRTC/SIP loops and runs synthetic conversational personas on local low-latency TTS models to tie up scam call centers.
• Dipteran: Direct mapping from the Drosophila melanogaster connectome into autonomous steering controllers inside SUMO traffic simulations.
• Project Impart: Android emergency paging app. It routes digitally signed alert packets through offline QR handshake mesh networks with zero cellular infrastructure.
• FluxBoard: Android keyboard engine. Runs swipe typing and next-word prediction entirely on-device with zero network permissions.

---

Source & Community

• Code: github.com/uncoalesced
• Web: uncoalesced.com
• Sponsorship: GitHub Sponsors