Work published on GitHub and Hugging Face
A portfolio of implementations published on GitHub and Hugging Face. The Hugging Face Spaces demos are playable without leaving this page. If you'd like to license or take over any of the code, that can be arranged too.
Core stack
$ git remote -v
origin git@github.com:you/your-product.git
$ transfer --repo my-awesome-repo --to you
✓ contract signed
✓ ownership transferred
→ 30 days of support includedScroll to explore
Portfolio
Pulled straight from the public GitHub and Hugging Face APIs and refreshed hourly. Where the same tool is published as code, a model and a demo, they are shown as one entry.
Data refreshes every hour.
Click any card under the "HF Spaces" tab and the app runs right here — no install, no sign-up, no leaving this page. Seeing how it behaves is faster than reading the source, so start there.
NagaYu/sumi
[feature-complete] Japanese PII detection that holds up on hard negatives — a 0.13B CPU token classifier, a rule layer, and reversible masking. Drops into Presidio in one line.
NagaYu/halfword
AAC word prediction optimised for time to utterance, not keystrokes saved. The divergence is driven by positional cost, not by how slow the access method is.
NagaYu/mondegreen
用語集を音韻的な硬い制約に変換するローカルASR訂正器 / Local Japanese ASR post-correction: a private glossary compiled into a hard phonetic constraint. Optimises for not breaking correct text.
NagaYu/molt
Elastic on-device inference: swap the LLM mid-sentence and carry the KV cache across. A generation that hits memory pressure migrates to a smaller model between two tokens instead of being reclaimed or restarted.
NagaYu/saccade
[closed] Always-on edge VLM research prototype: prediction-error gating + motion-compensated KV reuse + hard energy budgets
NagaYu/sieve
Deterministic RAG context auditor: tells you whether retrieval or generation failed, what context was wasted, and which chunks carry prompt injections or secrets. No model calls, no provider keys.
NagaYu/canary
Prompt version registry and deterministic regression gate for LLM changes — no model calls, no LLM judge
NagaYu/signpost
Rule checkers ask "is there a label?". Signpost asks "does the label mean anything?" — a small multilingual model for accessibility label quality, plus a constrained rewriter.
NagaYu/cue
Cue — a small multilingual model that scores subtitle break cost, plus a constraint-satisfying re-segmenter (SubER-first, guarantees normative compliance, strong on no-space languages).
NagaYu/rebar
Restore the heading hierarchy of Japanese documents from flat text, and chunk along it for better retrieval. CPU-only, layout-independent.
NagaYu/deference
Detect the direction of deference in Japanese honorifics (keigo), with the source citation. Where surface rules score 0%, this reaches 97%. 0.28B, CPU, 26ms.
NagaYu/tessera
Span-level, multi-label language identification. Recovers low-resource text that document-level LID discards. Runs behind GlotLID/OpenLID, not instead of them.
NagaYu/isotope
Untrusted data can inform an agent's action without authorizing it. Quantitative information-flow tracking inside the forward pass: attack success 0.44 -> 0.00 at 0.90 benign utility.
NagaYu/kvine
Branch-aware KV cache treating agent trajectories as a persistent copy-on-write tree — research prototype (PyTorch + HF transformers)
NagaYu/assay
Detecting model substitution with per-request internal receipts: signed hidden-state sketches verified at 0.63x a full recompute, with 100% coverage
NagaYu/downstep
Calibrated Japanese pitch-accent prediction for compounds that are not in the TTS dictionary. Abstains below a confidence threshold and defers to OpenJTalk, so it can be dropped into VOICEVOX / Style-Bert-VITS2 without making them worse.
NagaYu/flashback
Find the step your training run broke on, from 452 bytes/step. Delta-compressed branchable state history + all-step sketch index + consensus bisect.
NagaYu/ante
Accept contributions on evidence, not on authorship. An acceptance protocol that moves the burden of proof to the claimant — verifier, GitHub Action, Space, and a 42-PR benchmark.
NagaYu/parity
Certified vocabulary augmentation without continued pretraining: cut non-English token cost on open-weight models with a proven bound on behaviour drift.
NagaYu/stemma
Model provenance from weights alone: derivation direction, multi-parent merges and mixing ratios, recovered over HTTP Range reads without downloading checkpoints.
NagaYu/scribe-koyobun
Context-aware kanji/kana checker for Japanese official documents — judges the context-dependent usage that dictionaries and regex can't, and cites the normative clause. Rule layer + small token-classification model + Hugging Face Space/Model/Dataset.
NagaYu/dissent
Find formalization errors nobody is looking for: N independent formalizations of a contract clause, then an SMT search for the input where they disagree. Synthetic clauses only. Not legal advice.
NagaYu/litmus
Sound error bounds, symbolic GPU safety checks and targeted falsification for Triton kernels. 88% of planted bugs pass the standard fixed-shape allclose test; Litmus catches 100% with 0% false positives, on CPU.
NagaYu/groundhog
Share the world, resample the consequences: deterministic record/replay of exogenous randomness for branching agent RL. 29x lower advantage variance with no loss of counterfactual fidelity.
NagaYu/quorum
Rank-targeted nested sequential design for LLM evaluation: reach the same ranking conclusion for less, and see which comparisons the data never supported.
NagaYu/cairn
Leave, come back, same world — and edit it. An explicit world ledger makes video generation return-consistent (O(1) in absence duration) and editable.
NagaYu/llama-local-lab
Zero-Cost Llama Evaluation Studio — plan quantization, generate GGUF commands, model inference cost, and run Llama in your browser. No server, no API key.
NagaYu/phonebook
Japanese corporate names to katakana readings: a small CPU-friendly G2P with a copy mechanism, constrained decoding, calibrated confidence, and a known/unseen/hard/ambiguous split design built so its own claims can be refuted.
NagaYu/unquote
Email zone segmentation that never deletes body text. A 31M-parameter CPU model that tags every line as body/quote/attribution/signature/disclaimer/mobile-footer/auto-reply/forward-header, with a calibrated gate that keeps anything it is unsure about. 0.008% body loss vs talon's 4.18%.
NagaYu/crowd
Estimate how large a crowd a description hides in, then mask only what is needed. On-device anonymity-set estimation with exact counting, calibrated intervals, and a minimal-mask optimiser.
NagaYu/interlock
A shield that checks whether an agent's next tool call stays inside what the user actually asked for. Detects scope expansion that content-based shields cannot see.
NagaYu/ingot
Inverted generation: construct the label first, render the sentence second. Strictly-labelled, contamination-free synthetic data without a judge.
NagaYu/customs
Declare what leaves. Ration what it costs. Tiered LLM routing under a daily free quota, with data minimisation by query decomposition and machine-checkable egress records.
NagaYu/rubato
Predict the silence as a distribution, then decide when to speak. A model-agnostic turn-taking timing layer for spoken dialogue: calibrated silence-survival forecasting + optimal stopping under an asymmetric cost.
NagaYu/bleep
Sensitive speech regions without transcription
NagaYu/warden
Policy decisions + tamper-evident ledger for LLM traffic
NagaYu/ClaimCheck
Verify LLM answers against the context they came from
NagaYu/EchoCache
Semantic cache and cost observability layer for LLM APIs
NagaYu/sludge
Dark patterns in UI trees. Not a legality call.
NagaYu/PromptGate
Prompt injection, PII and secret detection in your browser
NagaYu/promptgate-rules
No description provided yet.
NagaYu/warden-policy-pack
No description provided yet.
NagaYu/claimcheck-rules
No description provided yet.
NagaYu/echocache-matcher
No description provided yet.
NagaYu/sludge-small
No description provided yet.
NagaYu/bleep-0.09b
No description provided yet.
NagaYu/scribe-usage-classifier
No description provided yet.
NagaYu/AdRadar
📡 Competitive Meta (Facebook) Ad Library radar — detect competitors' new creatives & long-running winners, alert Slack/Discord. Experimental.
NagaYu/IntentCraft
🪄 Turn your Swift code into Siri & Apple Intelligence (App Intents) integrations — 100% on-device, powered by Apple FoundationModels. CLI + native macOS app.
NagaYu/webgpu-radar
[single-device complete] Real WebGPU capability detection for browser AI — compute micro-benchmarks, community quirks DB, tiered recommendations with WASM fallback. Not just 'if (navigator.gpu)'.
NagaYu/smolvlm-coreml
SmolVLM-256M fully on Core ML + dependency-free Swift CLI with bit-exact HF preprocessing. Includes an empirical catalog of Core ML runtime constraints (stateful models, quantization, ANE).
NagaYu/pcc-verify
Verify your own Apple Intelligence requests against Apple's Private Cloud Compute transparency log
NagaYu/local-first-sync
Photo/file sync across all your nearby Apple devices at once — Multipeer Connectivity with async/await, per-peer progress, and SwiftData sync state. No cloud, no server, no account.
NagaYu/DualIntent
One Swift function → a Siri/Shortcuts App Intent and an MCP tool, generated at compile time
NagaYu/FoundationRadar
Measure what Apple's Foundation Models stack can actually do on this device — availability, real context size, real latency, PCC signals — instead of trusting a boolean.
NagaYu/hybrid-voice-separation
Educational demo: hybrid air + bone conduction voice separation (synthetic-data DSP pipeline)
NagaYu/audiencesync
Secure, in-memory Reverse ETL: extract high-value customers, SHA-256 hash their PII, and sync to Meta Custom Audiences + Google Customer Match. No CSV files, ever.
NagaYu/edge-gtm
Run server-side Google Tag Manager (sGTM) on Cloudflare Workers for $0-$5/mo, with first-party cookie rewriting that survives Safari ITP.
NagaYu/SwiftGuard
A fully on-device security & quality auditor for Swift code — powered by Apple's FoundationModels (on-device LLM / Private Cloud Compute). Your code never leaves your Mac.
NagaYu/plumb
Is the gate still vertical on what it skips? An audit layer for gated always-on egocentric perception (EPIC, EgoTrigger, Ego-METAS): skipped-part degradation, drift, and a bit-identical zero-trust guarantee. Synthetic-only; negative results included.
NagaYu/clearance
Structural isolation and revocation propagation for shared vector indexes. Your ACL filter protects the index, not the summary you built from it yesterday. Research prototype, synthetic data only.
NagaYu/dendro
Prove a document existed before a date from independent archives, instead of guessing whether it looks generated.
NagaYu/provably-local-mobile
Prove in your CI that an on-device feature needs no network. Android (Kotlin) + iOS (Swift) testing libraries. MIT.
NagaYu/local-first-llm-react-native
On-device vs. cloud routing policy for React Native LLM apps. The RN sibling of local-first-llm, wrapping react-native-executorch.
NagaYu/executorch-matrix
Compare ExecuTorch hardware backends and quantization levels for your model on your actual device, with a reasoned recommendation. Honest about what needs real hardware.
NagaYu/local-first-llm-android
Provider-agnostic, per-request on-device/cloud LLM routing for native Android (AICore + any cloud). Kotlin sibling of local-first-llm.
NagaYu/aicore-radar
Measure what Android's on-device AI (AICore / Gemini Nano via ML Kit GenAI) can actually do on this device — typed availability, quota/foreground error wrapping, latency benchmark, measured context window. No telemetry.
NagaYu/tier-matrix
Mock every capability tier your on-device AI app can fall back to and re-run a test through each one — not just the happy path.
NagaYu/local-first-audit
Statically scans a JS/TS codebase for LLM API call sites and classifies each one's on-device migration suitability
NagaYu/webmcp-sidecar
Bring your own model to WebMCP-enabled sites: a side panel AI assistant that discovers and calls tools a page has registered via WebMCP.
NagaYu/local-first-rag-apple
On-device semantic search for Apple platforms that also powers system Spotlight and Siri/Shortcuts — NLEmbedding + SwiftData + Core Spotlight + App Intents, one API surface.
This is a selection — there are 99 public repositories in total.
See them all on GitHubLicensing
Before getting in touch, it's worth running the demo for whatever you're interested in.
Stated up front so there are no surprises later.
These are the standard terms. Where an individual contract differs, the contract governs.
Contact
For transferring or licensing published code. Every message gets a reply as soon as I can get to it.