Projects and validation

Spring Boot Agent Gateway

What stands out：Use tool traces to diagnose invalid arguments, blocked tool calls, and timeouts in a Java integration.

My contribution

Built with Java 21 / Spring Boot 3: an Agent Chat API, bounded loop, allowlisted Tool Registry, JSON Schema validation, tool traces, session storage, and minimal RAG.

Tool management

Registers knowledge_search and structured read_only_query ; execution is controlled through an allowlist, JSON Schema, timeouts, and bounded retries.

Validation

Compiled on JDK 21.0.12 and Maven 3.9.16. Six tests across four classes passed, covering tool execution, RAG chunking/retrieval, and the application Agent loop.

How to run

mvn test ; local checks use H2 and an in-memory adapter. MySQL, Redis, and online-model integrations require further verification.

Extension points

The default HashEmbeddingPort validates the service flow. The EmbeddingPort interface supports replacement and retrieval evaluation on specific datasets.

Text2SQL / Data Agent

What stands out：Complete composite join keys along trusted paths, and stop when the path is ambiguous or unauthorized.

User task

Turn natural-language data requests into a flow with clarification and validation, while constraining generated SQL before execution.

My contribution

Built on a shared training-program LangGraph workflow, maintained the clarification, schema retrieval, SQL generation, and execution flow, and tightened SQL validation and read-only safeguards.

Verification results

The September 10, 2026 local safety checks passed 11 offline regressions and two MySQL permission tests. Natural-language SQL generation accuracy and end-to-end task completion remain unverified.

Data environment

The Agent prototype uses a local MySQL sample database. The separate schema evaluator uses synthetic SQLite data.

Open-source contribution

Contributed Chinese and English explanations of schema selection and join paths to awesome-ai-roadmap. After maintainer review and revisions, PR #17 was merged on September 29, 2026.

Follow-up evaluation

Built a Python / SQLite evaluator with 8 synthetic tables, 9 test inputs, and 34 automated tests for join-key completion, permission constraints, path ambiguity, and insufficient information.

Verification method

Compare column recall before / after completion, complete query policies, and fixed reference SQL results. Require clarification for ambiguous joins and retain refusal branches when no path or permission is available.

Collaboration Radar

What stands out：Separate proposals from confirmed decisions, trace tasks to their sources, and leave unstated deadlines open.

Public sources

ClawHub： Chinese edition 、 English edition 。

What it does

I implemented seven-module collaboration cards and a deterministic processing layer for source tracking, PII masking, cross-round aggregation, schema validation, conflict detection, and CSV / Feishu Base / Notion field-mapped exports.

Self-test coverage

Adversarial inputs cover CSV formula injection E4, SVG XSS T5, Markdown table injection M1–M5, malformed JSON, type mismatches, and missing files.

Implementation

Uses a fixed schema, deterministic processing, and selftest to turn meeting notes and group chats into structured status cards.

Use cases

Organizing collaboration materials, summarizing status, and exporting structured records.

Read the full case: input → status cards → source review → export ↗

WeRead AI Brain

What stands out：Compare ideas across books around one question, check passages, and remove unsupported links before export.

Public sources

GitHub：yamyeed/weread-ai-brain。

What it does

Combines WeRead data, a visual dashboard, AI analysis, and Markdown knowledge export. Skill documentation includes dependencies, commands, and environment variables.

Prerequisites

Users provide an API key through local environment variables. The public repository contains no personal keys.

Usage

Available as an open-source repository and Skill configuration. The public cross-book walkthrough uses manually written interpretations for source review, correction, and export, without connecting to a personal WeRead account.

Annual-report sentiment analysis delivery

What stands out：Extract topic-specific sentences from annual reports and export sentence-level results in a Windows workflow the client could run.

Business problem

The client needed to filter topic-specific statements and identify sentiment across annual reports. Manual review was slow, and nontechnical users could not run model scripts directly.

Requirements and iteration

I broke an open-ended request into text processing, topic filtering, sentiment classification, validation, and export. I delivered a single-company version first, then expanded batch inputs and XLSX fields based on actual needs.

Deliverables

Packaged BERT-based sentence sentiment analysis in a Windows GUI, with drag-and-drop annual reports or ZIP batches and Excel / CSV exports.

Runtime and handoff

Repackaged the environment to address missing dependencies, GUI startup failures, slow batches, and network issues, then added a demonstration and operating guide.

Delivery outcome

The client completed batch processing and structured exports using the guide, then accepted and paid for the project.

Product scope

The tool supports investment research and information review; final investment decisions remain with the user.

Agent Task Evaluation

Check task completion and trace failures through tool calls, file handling, and saved deliverables.

User task

Compound tasks require tools, file processing, and saved results. Evaluation checks the actual artifacts from each step against the task requirements.

My contribution

Recorded tool calls, code execution, data analysis, and saved deliverables across five model aliases, three compound task types, and two execution environments.

Verification

Compared task requirements, execution records, and screenshots of evaluation artifacts. Located failures in task understanding, tool inputs, execution, saving, or supporting evidence.

Findings

Documented context loss, interrupted tool chains, inconsistent artifact formats, and conclusions without supporting data, informing checks, human confirmation, and error recovery.

Material scope

This page summarizes methods and failure patterns. Original task materials are confidential, and a complete public redacted package remains unavailable. The records support observations within the specific tasks.

RAG and MCP tool-protocol validation

What stands out：Evaluate retrieval, tool interoperability, and application permissions as separate concerns.

What it does

Built parallel LangChain and LlamaIndex pipelines to compare chunking, dense + BM25 retrieval, and RRF fusion, with Chroma persistence.

MCP validation

Implemented a minimal JSON-RPC 2.0 over stdio MCP server using Python's standard library, registering two tools. A custom client completed initialize、tools/list and two tools/call smoke tests.

Engineering decisions

The framework comparison checks pipeline composition while retrieval quality depends on chunking, retrieval, and ranking. MCP validates the tool protocol; permissions, audit, and human confirmation remain application responsibilities.

Test record

Covers protocol initialization, tool discovery, and calls to two tools through a custom client.

Expense pre-check: AI-assisted evidence review

What stands out：Distinguish missing evidence from a policy violation, and explain what to supply before rule review.

Scope

Checks evidence completeness before expense-rule review. Missing attachments, conflicting invoice fields, or insufficient exception evidence are routed to FLAG with requests for supporting material or human verification. Final approval remains human.

A correction

A rule review found that the initial logic grouped missing attachments with confirmed policy violations under REJECT . The revised rule treats insufficient evidence as pending verification. Test T10 confirms that missing attachments do not trigger automatic rejection.

Validation scope

12/12 routing tests passed. Accuracy on 30 labeled samples was 63.33%, below the predefined 80% threshold; human review remains in scope.

Demonstrated capability

Narrow high-risk scope, distinguish missing evidence from violations, use evaluation thresholds to decide expansion, and preserve actionable next steps for reviewers.

Scientific Nutrition Advisor

What stands out：Suggest takeaway options around budget and preferences, clarify missing details, and flag situations that need professional guidance.

Public sources

Chinese edition 、 English edition 、 iFLYTEK award announcement 。

Download statistics

Cumulative downloads for each language edition are available on its public ClawHub page, using the platform's distribution metric.

Health guidance

Product guidance includes referral paths for special populations, medical conditions, and medical advice. The tool supports health management.

AI Reading Coach

What stands out：Change the work goal to see a different reading priority for the same technical book.

Product task

Turn one-off AI questions into a stateful, trackable reading workflow: tailor a book map to the reader's background, difficulties, and goals; guide each chapter; then check understanding through application questions.

My contribution

Specified data models, chapter parsing, reading batches, progress states, generation constraints, fallback behavior, and acceptance criteria in structured requirements readable by AI tools, then developed the UI prototype.

Product boundaries

The HTTP Agent interface supports SSE, NDJSON, JSON, and plain text. The public demo uses a local streaming engine. The developer panel exposes the generation source, estimated tokens, and timing.

Validation

The November 3, 2026 local preview passed 40 tests, type checking, and the build. Tests found and helped fix a bug that ignored supplied chapter state when creating a book.

Current status

The open-source application includes a bookshelf, chapter progress, and an Agent integration interface. This offline demo matches goal terms to chapter titles. Generated content and questions demonstrate the flow; target-user outcomes remain unverified.

Media Extract

What stands out：Use existing text layers or subtitles first, keep source references, and retry only unfinished items.

Routes materials through text layers, embedded subtitles, OCR, or speech recognition, producing readable text, structured metadata, and source artifacts.

Read the full case ↗

PersonalOS Tool-Calling Agent

User task

Let users ask about knowledge and articles, with an assistant invoking retrieval tools and retaining reviewable conversations.

Choice & cost

I bounded visible tools, history, and loop rounds to control execution. Keyword-based selection may omit a tool and trimmed history may lose context, creating a tradeoff between task coverage and execution budget.

Validation & next step

Existing records support inspection of tool contracts, conversation storage, and execution limits. Next: test multi-turn follow-ups and retrieval failures for missing information and recovery; runtime success alone does not establish answer quality.