Language Model Loss Captures Mutual Intelligibility Gradients in Turkic Languages
Moldir Baidildinova, Shiva Upadhye, Austin Wagner (UC Irvine)
- Accepted at HSP 2026 (39th Annual Conference on Human Sentence Processing, MIT) and Tu+11 (11th Workshop on Turkic and Languages in Contact with Turkic, MIT)
- Character-level LSTM trained on 6 Turkic languages + Finnish (control), all transliterated to broad IPA
- Built on Turkic API — rules-based IPA transliteration with phonetic rules cited from peer-reviewed research
Plant Metabolomics & BVOCs
- Research on biogenic volatile organic compounds from tree leaf and root tissues across 9 California field sites
- GC-MS data processing pipeline: Tree Bot — peak detection, co-elution analysis, compound classification
- Interactive Metabolomics Dashboard — Welch's t-test with FDR correction, D3.js visualizations
All services live in my API monorepo — 13 microservices, 21 shared libraries, FastAPI + Redis + RQ architecture.
All Services (13) & Shared Libraries (21)
See the full API monorepo for all services and libraries including Data Bank, Music Wrapped, Transcript API, Handwriting AI, QR API, GitHub Stats, Opportunity Radar, Procart API, and 21 shared libraries covering ML, NLP, workers, persistence, and more.
Interactive data dashboards at austinwagner.org — see the Dashboards repo.
Multi-tenant AI operations platform (private repo — access on request)
30+ MCP servers powering document ingestion, meeting parsing, and knowledge synthesis. Custom 4-layer architecture enforcement — workspace guard rules + architecture tests + boot-time checks + MCP-tool rejection — that lint invariants CI-time so drift can't ship. Wiki-in-DB with primary-source audit chain (RLS-sliced tenant scoping, audit trail on writes). Tenant provisioning via canonical-writer pattern with policy-enforced sandbox namespaces. Deterministic parser fleet for meeting sources (5 platforms, 34+ agencies) with codegen-emitted migrations from a typed source registry. Semantic search via Nomic Embed v1 + HNSW (hnswlib-node). TypeScript + Python, Postgres 16 with RLS, Keycloak OAuth 2.1, Docker Compose orchestration.
Personal research wiki (private — architecture write-up on request)
559 atomic pages across 21 topic hubs (nursing quality improvement, mass surveillance & civil liberties, computational linguistics, atmospheric chemistry, QFT, metabolomics, others). SHA256-verified primary-source audit chain via local_pdf: frontmatter + a files/wiki archive; every claim is mechanically re-verifiable against a locally-archived PDF. Semantic search via Nomic Embed + HNSW. Cross-domain synthesis layer — pages can link into multiple hubs (polyhierarchy).
Contributor to microsoft/LightGBM — 9 PRs merged across Python type safety, C++ networking fixes, and test coverage.
| PR | Description |
|---|---|
| #7137 | Fix socket timeout on POSIX systems (wrong type for setsockopt) |
| #7131 | Fix test_numeric_split_direction to test all parameter combinations |
| #7133 | Add test for Booster.rollback_one_iter() |
| #7130 | Fix numpy integer cast in plot_importance |
| #7115–7119 | Type annotations — TypeGuard, Literal, DTypeLike, return types for sklearn predict |


