_{RENCHAO WU / WIDECHAOS}

智灯已灭余空烬，犹自光明照十方
_{"From the ashes a fire shall be woken, a light from the shadows shall spring." — J.R.R. Tolkien}

_{NOW / Voice agents & realtime systems}

./manifesto

Small tools for inspecting text, datasets and model results. Pick a task to explore.

Explore all experiments

• Bilingual Router LoRA — Test a small model that turns bilingual requests into structured actions.
• Evidence Agent — Inspect an agent’s answer through its tool calls and source traces.
• JSON Wire Browser Lab — Inspect how decoding and JSON round-trips change values and bytes.
• Offline Evidence Finder — Find exact quotes in English or Chinese text, with source offsets.
• Tiny Hash Intent — Try a small English intent classifier and inspect its retained errors.
• Split Boundary Audit · Group Planner — Plan train, validation and test splits while keeping declared groups together.
• Annotation Pair Review — Compare two annotation files and manually resolve disagreements.
• Local Label Queue — Label text while keeping abstentions and unfinished records separate.
• Prediction Metric Desk — Audit saved predictions, missing outputs and the denominators behind scores.
• Bike Hour Forecast — Estimate the next historical rental hour and compare two simple baselines locally.

Methods, results & limits

• Bilingual Router LoRA: Bilingual Router: strict all-field accuracy 7.6% to 47.2%, JSON validity 98.6% to 95.5%; 576 synthetic tests, one run. Read results and failures. Qwen2.5-0.5B-Instruct, base → adapter, 576 frozen synthetic tests. One seed and run, with the schema supplied to both models. Strict all-field accuracy improved while strict JSON validity fell; unseen-operation failures remain visible. These are experimental results, not production performance. Model & reproduction.
• Evidence Agent: Evidence Agent: historical local 14B evaluation 9 of 20; live 4B demo has no equivalent verified score. Traces do not guarantee correctness. Read evaluation and failures. the 9/20 result belongs to the documented local Qwen3-14B run on October 3, 2026. The live 4B demo has no equivalent verified score. Tool traces support inspection, not a correctness guarantee. Free GPU quota can interrupt inference; labeled tool walkthroughs remain separate. Demo & setup.
• JSON Wire Browser Lab: A parsed value can survive a round-trip: Bytes ≠ values. Decoding, parsing and encoding differ; Not a semantic correctness guarantee. A browser-local workbench separates UTF-8 decoding, JSON parsing and re-encoding, with fixed byte fixtures and a frozen Python reference. A large integer can change even when the parsed value round-trips. These checks do not prove semantic correctness or sanitize untrusted input. Exported reports include input bytes and runtime details; review them before sharing. Public lab & source.
• Offline Evidence Finder: Unanswerable queries returned a hit: 1 / 4. Lexical matching, not semantic support; Hand-picked development queries. Browser-local English and Chinese lexical retrieval exports exact quotes, UTF-16 offsets and source text. After a disclosed sample-extraction correction, one of four unanswerable queries still matched an unrelated excerpt. These hand-picked development queries are not a blind retrieval benchmark; a location match does not establish truth, answerability or semantic support. No model inference or training. Public lab & source · English write-up · 中文说明.
• Tiny Hash Intent: Test errors · One fit, four views: 301 / 3,080. English closed set · No OOD rejection; Int8 storage is not a speed benchmark. One from-scratch fit on BANKING77 is evaluated as float/int8 storage with raw/calibrated probabilities: four representations, not four trained models. All 301 errors on the official 3,080-row test set are retained; the two quantization-changed predictions are wrong before and after. This is a previously seen benchmark, not a historical blind test. The English 77-class closed set always selects a known class, with no out-of-domain rejection. The browser dequantizes int8 storage for floating-point inference; no integer-kernel or speed-up claim. Public model demo & source · English write-up · 中文说明.
• Split Boundary Audit · Group Planner: Actual split · 36-record synthetic example: 30 / 3 / 3. Greedy planning is not optimal; No independence or semantic guarantee. A local browser planner keeps caller-declared groups and groups connected by exact text together across train, validation and test, then exports the full plan and assigned records for the included Python auditor. It preserves every record, conflicting labels and achieved-ratio deviations. The 36-record synthetic example has 12 declared groups and 10 connected units; an 80/10/10 target yields 30/3/3 records, with one normalization candidate and one conflicting-label set retained for review. Normalized matches are candidates only. A retained counterexample shows the single-pass greedy planner is not globally optimal. Group provenance is supplied by the caller; a CLEAR audit does not prove statistical independence or absence of semantic leakage. Original Apache-2.0 code and synthetic examples; no model or external dataset. Public planner, source & audit · English write-up · 中文说明.
• Annotation Pair Review: Partial synthetic example · Both labelled: 100% · 1 / 12. Agreement is not accuracy; No autosave · Export to keep work. A browser-local workbench shows conditional label agreement beside coverage, keeping disagreements, abstentions and unfinished records visible for manual final decisions. The partial synthetic example reaches 100% agreement with only 1 of 12 records labelled by both roles; the default example has 4/12 coverage and 2/4 identical labels (50%). These are scripted role annotations, not independent human annotators, accuracy or gold labels. Full inventory, taxonomy and record order must match; no intersection-only merging or automatic adjudication. The tab does not autosave: refresh loses unsaved work. Downloaded snapshots can restore a review and contain full original texts and notes; review them before sharing. Checksums establish consistency, not authenticity. No model, training or measured adoption. Public workbench, source & reproduction · English write-up · 中文说明.
• Local Label Queue: 12 synthetic records · All abstained: 100% · 0 labels. Review progress is not label quality; No autosave · Download a checkpoint. A browser-local manual text-labeling queue keeps labelled, abstained and unfinished records separate and preserves every source text, id, group and annotation in exports. In the synthetic all-abstained example, all 12 records are reviewed but zero labels are assigned. The running demo instead has 2 labelled, 1 abstained and 9 unfinished records; revision 5 counts edits, not reviewed records. These authored examples have no gold labels or independently verified human annotators. No model, training, automatic quality check or measured adoption. There is no autosave: refresh loses unsaved work. Download a checkpoint first; restoration requires a verified preview and explicit full-session replacement, with no merge. Exports contain full source texts and annotations, so review before sharing. Hashes establish consistency, not signatures, encryption or authenticity. Public queue, source & reproduction · English write-up · 中文说明.
• Prediction Metric Desk: Conditional accuracy · Known-label coverage: 75% · 4 / 7. Synthetic example · Different denominators; Completion bounds are not confidence limits. A browser-local classification-result audit keeps the full inventory beside missing predictions, abstentions, duplicate IDs and invalid rows. In the eight-record synthetic fault example, conditional accuracy is 3/4 (75%), known-label coverage is 4/7 and total decision coverage is 5/8: different denominators, not interchangeable scores. Hypothetical completion bounds are not forecasts or confidence intervals. Saved BANKING77 predictions are recalculated, not a new model, inference, calibration run or blind benchmark; seven normalized train/test overlaps remain. No automatic repair, renormalization, clipping or invented gold labels. Full source texts and predictions appear in exports; review before sharing. Memory-only, no autosave or checkpoint restoration; refresh clears work. Hash consistency does not prove authenticity or gold-label quality. Code and synthetic fixtures are Apache-2.0; BANKING77 original and derived data retain CC BY 4.0 attribution. The standalone App passed the owner release gate; Hub embedding remained unverified in the original Edge context, with no established cause. Standalone audit desk · HF source, methods & licenses.
• Bike Hour Forecast: Historical test MAE · 4,258 eligible targets: 33.85 vs 86.91 / 80.94. 118 test targets excluded for incomplete history; System-level 2011–2012 counts; no current-city or station availability validation. One 1,185-parameter network trained for 40 CPU epochs; epoch 20 was selected using validation only. The browser runs saved weights on 24 consecutive hourly counts and exports all inputs and predictions beside two untrained baselines. On the same 4,258 eligible test targets, MAE is 33.8476 versus 86.9054 and 80.9385; 118 of 4,376 test targets were omitted for incomplete history. The largest retained error predicts 509.01 against truth 114. Source data have 165 absent nominal hours, with no zero-fill or UTC/DST reconstruction. Rolling forecasts may use observed earlier validation/test counts, windows overlap, and source data were seen before the protocol. No historical-blind, uncertainty, stronger-baseline, current-city, station-availability, production or adoption claim. Today’s release runs saved weights; it is not new training. Code is Apache-2.0; UCI data and derived weights/results are CC BY 4.0 with full attribution. Public HF source & methods · Exact release.

Contribution details →

AI & Agents
 LangGraph ·  LangChain ·  Hugging Face ·  Transformers ·  Ollama ·  RAG ·  MCP

Machine Learning & Data
 PyTorch ·  TensorFlow ·  scikit-learn ·  Pandas ·  NumPy ·  Matplotlib

Voice & Realtime
 LiveKit ·  FunASR ·  faster-whisper ·  Azure Speech ·  WebRTC ·  SIP / FreeSWITCH

Backend
 Python ·  Django ·  FastAPI ·  Flask ·  Go ·  Pydantic

Databases & Retrieval
 PostgreSQL ·  MySQL ·  SQLite ·  Redis ·  SQLAlchemy ·  Vector Search

Web Crawling & Automation
 Scrapy ·  Beautiful Soup ·  Playwright ·  Selenium

Cloud & Storage
 Docker ·  Kubernetes ·  Linux ·  Nginx ·  OpenStack ·  Ceph ·  Bash

Frontend
 TypeScript ·  JavaScript ·  Vite ·  Astro ·  HTML5 ·  CSS

Testing & Delivery
 pytest ·  Git ·  GitHub Actions

Repository languages

Verify learning credentials

• Databricks Academy · Generative AI Fundamentals
• Databricks Academy · AI Agent Fundamentals
• Databricks Academy · AI Security Fundamentals
• Hugging Face · Fundamentals of Agents · Unit 1 certificate
• Hugging Face · Agents Course certificate
• Alibaba Cloud · Apsara Clouder · Generative AI
• Cognitive Class · Prompt Engineering for Everyone certificate
• Cognitive Class · Build Your Own Chatbot certificate
• Cognitive Class · Fundamentals of AI certificate
• Cognitive Class · Machine Learning with Python certificate
• IBM SkillsBuild · Artificial Intelligence Fundamentals certificate
• Kaggle Learn · Intro to Programming certificate
• Kaggle Learn · Python certificate
• Kaggle Learn · Intro to AI Ethics certificate
• Kaggle Learn · Intro to Machine Learning certificate
• Kaggle Learn · Pandas certificate
• Kaggle Learn · Data Cleaning certificate
• Kaggle Learn · Intermediate Machine Learning certificate
• Kaggle Learn · Intro to SQL certificate
• Kaggle Learn · Data Visualization certificate
• Kaggle Learn · Advanced SQL certificate
• Kaggle Learn · Intro to Deep Learning certificate
• Kaggle Learn · Feature Engineering certificate
• Kaggle Learn · Machine Learning Explainability certificate
• Kaggle Learn · Time Series certificate
• Kaggle Learn · Computer Vision certificate
• Kaggle Learn · Geospatial Analysis certificate
• Kaggle Learn · Intro to Game AI and Reinforcement Learning certificate

_{Kaggle · DEV}

_{WeRead · Reading milestones}

More reading badges

Event mementos · Virtual sticker books

_{Hacktoberfest 2026 · Virtual sticker book
Official account dashboard (sign-in required)}

_{Per aspera ad astra.}

_{STILL BUILDING // WIDECHAOS}