Projects | Pengbin Li
Six projects that show how I approach agent orchestration, RAG quality, safety, evaluation, and applied deep learning.
Trip Planner · AI Travel Planning System
Jun 2026
A Chinese-language AI travel planning product that connects fast itinerary generation, deep destination research, conversational editing, maps and weather enrichment, history management, and Markdown/PDF export into one continuous workflow.
Stack： Vue 3 · TypeScript · FastAPI · LangGraph · ChromaDB · Redis · SQLite · RAG
Highlights
- Built a Destination Intelligence agent with chapter planning, web research, evidence synthesis, reflection, and source tracing. Across 8 real reports it averaged 5 sections, about 103 sources, 84 unique URLs, and 33 minutes, with 8/8 saved-result completion.
- Created a Report Itinerary agent that converts long-form Markdown and historical reports into frontend-ready itineraries through section chunking, structured extraction, quality gates, cache fingerprints, and resumable execution.
- Designed an Ask / Search / Research / Update LangGraph chatbot with itinerary Q&A, live destination queries, multi-step research, comparison, risk checks, and natural-language itinerary updates.
- Implemented scoped long-term memory in SQLite and a Chroma-to-keyword fallback path for semantic, episodic, and perceptual memory operations.
- Improved offline RAG Top-1 accuracy from 53.3% to 93.3%, Top-K to 100%, and reduced cross-destination contamination from 17 chunks to zero.
- Added Redis-backed caching for RAG, maps, geocoding, route estimation, and weather, with graceful degradation when Redis is unavailable.
- Built a tau3-style evaluation harness. The latest real-LLM, real-tool run covered 116 cases and 126 turns with 100% case pass, turn pass, intent accuracy, tool-action match, and high-risk pass.
- Reduced first feedback latency from 0.4145s to 0.0052s in slow-search simulation through SSE streaming, a 98.7% improvement.
Enterprise Java AI Agent Platform Enhancement
Jun – Jul 2026
Extended java-up-up/super-agent with production-oriented capabilities for tool extensibility, RAG evaluation, retrieval closure, security, observability, and citation traceability.
Stack： Java · Spring AI Alibaba · RAG · Neo4j · Redis · Kafka · pgvector · Elasticsearch
Highlights
- Designed a pluggable enterprise Tool Registry around Spring AI Alibaba ReactAgent, including MCP adapters, enable/disable controls, timeout, quotas, audit, and fallback behavior.
- Built an offline RAG evaluation suite using historical traces and human golden sets to measure Recall@K, MRR, nDCG, citation hit rate, and latency across retrieval, reranking, and prompt variants.
- Added bounded iterative retrieval: evidence sufficiency check → query completion → second retrieval → evidence merge, limited to one retry and verified through traces and offline evaluation.
- Introduced prompt-injection defenses and citation-consistency checks for RAG and web-search evidence, isolating external content from executable instructions and enabling answer-to-source backtracking.
MyAgent · Defensive Security Assessment Agent
May – Jun 2026
A PI Agent-based CLI/TUI system enhanced with authorized defensive-security subagents, scoped execution, security memory, multi-stage planning, and report generation.
Stack： TypeScript · Node.js · PI Agent · CLI/TUI · Multi-Agent · Security Guardrails · Vitest
Highlights
- Injected security-mode prompts, tools, and contextual memory when security intent is detected at session startup.
- Required explicit target assets, allowed actions, authorization purpose, and expiry before active checks can run.
- Split work into asset discovery, attack-surface analysis, PoC research, risk assessment, remediation, and reporting, controlled by step and iteration plans.
- Added Asset Discovery, Attack Surface Analysis, and PoC Research subagents for baselining, exposure analysis, and vulnerability-source normalization.
- Expanded tools for passive intelligence, constrained crawling, CVE/NVD/Exploit-DB/MITRE normalization, security memory, web analysis, and Markdown/JSON reports.
- Covered intent recognition, authorization expiry, tool registration, memory persistence, delegation, human confirmation, and report generation with Vitest.
Paper Humanizer Skill
Jan 2026
An open-source bilingual academic-writing tool that reduces formulaic AI writing patterns while preserving facts, numbers, citations, conclusions, and domain terminology.
Stack： Claude Code Skill · Python CLI · Prompt Engineering
Highlights
- Designed bilingual academic-revision prompts, user templates, tone configuration, and a standardized four-part output.
- Created an AI-pattern phrase blacklist to reduce templated wording without flattening domain-specific style.
- Added strict factual-preservation constraints to prevent revision-induced distortion.
- Published a Python CLI and example data for batch processing and standardized prompt generation; reached 96 GitHub stars at CV publication time.
AutoPentest-XL · Autonomous Multi-Agent Pentest Platform
Jan – Feb 2026
An authorized security-testing platform using a PDEF loop—Perception, Decision, Execution, Feedback—to coordinate reconnaissance, vulnerability analysis, exploitation, privilege escalation, and post-failure reflection.
Stack： Python · LangGraph · LangChain · ChromaDB · Neo4j · Docker · Scapy
Highlights
- Modeled Commander, Scanner, Exploiter, Privesc, and Reviewer roles in LangGraph with reflection retries, context compression, dynamic routing, and human takeover.
- Combined ChromaDB for tool and vulnerability knowledge with Neo4j for topology, attack paths, and long-term memory, supporting CVE retrieval and GTFOBins mapping.
- Implemented CIDR boundary checks, dangerous-command blocking, reviewer approval, and Docker isolation so actions remain within authorized scope.
- Added faithfulness grading, OpenTelemetry, and Prometheus monitoring for token cost, execution time, and exploit success rate.
Sendai Satellite Image Semantic Segmentation
Aug – Sep 2024
Applied ViT and LoRA fine-tuning to high-resolution satellite imagery from Sendai, identifying roads, vegetation, buildings, and vehicles for urban planning and post-disaster emergency analysis.
Stack： Python · PyTorch · Vision Transformer · LoRA · OpenCV
Highlights
- Built the data preprocessing, annotation, augmentation, fine-tuning, and semantic-segmentation workflow for a practical remote-sensing scenario.
Start a conversation
If you are building in AI agents, RAG, AI security, or applied AI, you can review my full CV or contact me directly.