Vertical AI Agents across Government, Finance, Manufacturing, Consumer & Content — browse by industry
Government · SOE Regulation
Built for government and state-owned-asset supervision, it provides intelligent review, knowledge-base management, and a risk-label system, using multi-agent collaboration to assist scientific decision-making and full-chain risk control. This is our flagship production product (imported from the original soe_board_agent project).
Finance · Credit Approval
A unified view covering enterprise analysis, industry reference, risk alerts, and conclusion output — keeping credit analysis continuous, clear, and reviewable from upload to conclusion. The workbench covers the full flow of document identification, data extraction, metric review, information enrichment, relation mapping, and conclusion output, and delivers three-dimensional (legal, financial, industry) expert assessments plus a directly routable approval report.
Products · Finance Intelligence
Built for finance departments in precision-manufacturing and other large enterprises, the system uses the paradigm of "agent pre-audit + risk tiering + human adjudication": AI takes over the four repeated steps — invoice verification, duplicate detection, standard matching, and image check — while the finance manager keeps the final adjudication right. We ran a 6-day PoC for a 480-person manufacturing client and cleared all 300 pending reimbursements within 5 business days: manual review workload dropped 76%, average handling time fell from 15 to 4 minutes, and every abnormal judgment came with evidence — 100% explainable.
Manufacturing · Industrial Software
Automotive harness drawings carry large volumes of non-standard, cross-platform, strongly-connected engineering information; going from drawing to production board diagram has long relied on manual reading, organizing, and drafting. We have built an "agent exploration pre-processing" prototype — it reads raw schematics directly, automatically identifies time candidates, geometric relations, label bounding boxes, and upstream/downstream connections, and outputs a structured Excel BOM. Understanding harness drawings is one of the key hard problems in industrial-drawing AI-ization.
Products & Solutions · Industrial Software · Connector Data
A connector & terminal database system built for the automotive wiring-harness industry: covering 13 mainstream manufacturers, 3,875 series, 24,776 connectors, 16,519 terminals, and 1.97M+ mate relationships — paired with a dark-themed collection monitor for live progress. This project is a sub data module under "Wiring Harness Drawing & Board Generation", and itself is built on the "RPA Agent Skills System" to do cross-vendor structured crawling, ingestion, and version maintenance — collapsing connector data scattered across vendor sites, PDF catalogs, and Excel sheets into a unified data asset for engineers.
New Consumer · Food Tech
XM7 is an AI reverse-formula design system (digital-twin R&D lab) for tea and new-style beverage brands: starting from commercial goals, it reverse-derives flavor structure and ingredient formulas, then auto-synthesizes, validates, and produces a sample-ready commercial plan. It moves F&B R&D from "blind trial-and-error" to "deterministic innovation."
Content · Writing Aid
Centered on the "premise reverse-derivation" paradigm (protagonist—situation—book—genre—web), it decomposes long-form novel creation into four phases — original design / structure building / generative production / review — across 30 ordered steps, covering inspiration capture, theme setting, world framework, protagonist building, full-book outline, foreshadowing, emotional-curve plotting, and review. Integrated with Ollama online LLMs, it enables AI online writing and fine-grained control of every line · 200 segments.
Content · Story Creation
An industrialized creation system for high-completion-rate suspense / twist short stories on paid-story platforms like Yanxuan. Its core paradigm is "5 Foundation files + 11 gates + 9-step pipeline": it hard-codes front-loaded constraints across the plot_summary → world / characters / outline / voice / canon five phases, then drives the story from setting to a publishable draft through a 9-step pipeline — Foundation alignment → plot generation → initial generation → scene generation → editorial review → reader feedback → author revision → logic verification → proofreading gate → style rewrite. With 11 Foundation gates (F1–F11), a 10-question adversarial test, and multi-author style-fingerprint rewriting, it upgrades "human reviewing human" into "machine gates, human decides."
Product & Solution · AI Content Generation · Visual Science
Comic Builder is an AI tool that industrializes the production of multi-panel science-popularization comics from a one-sentence topic. Built on the baoyu-comic skill, it first uses an LLM to parse the requirement into content analysis + 6-segment storyboard, then generates an image prompt per panel, calls the image-generation API panel-by-panel, and finally stitches all panels into a single comic. The full progress is pushed to the frontend via SSE in real time, with a 6-panel typical task completing in 1-3 minutes. Designed to turn abstract science / engineering / business concepts into publishable visual content.
Customer Service · Sales Agent
Turn "sales/service calls" into an autonomous voice agent: on the front end, an ESP32 add-on takes over any phone's call audio via Bluetooth using the HFP protocol with zero friction; on the back end, it pushes via SIP/WebRTC to the Rust-based active-call voice engine, where an LLM drives real-time dialogue, sales conversion, and auto-answer. Hardware BOM under 20 CNY, with offline ASR/TTS and cloud dual engines — privacy-compliant and plug-and-play.
Reverse Engineering · Mobile Security
Using Maersk's official app as a sample, this fully demonstrates our Android reverse-engineering capability: static decompilation, Frida dynamic hooking, SSL certificate-pinning bypass, traffic-redirection capture, device certificate/key extraction, TLS-fingerprint replication (OkHttp/Android Chrome), and Akamai protection bypass — ultimately restoring a closed mobile interface into a standard, reusable API automation chain.
Experiments & Research · Device Control
A multi-device collaborative control system that combines LLM task orchestration with a 20-¥ self-built ESP32 ADB group-control router. The local proxy / pairing router carries all multi-device pairing info on the go, so a fleet of phones can be assembled into an ad-hoc group-control workbench on any network. An LLM converts natural-language instructions into executable click / input sequences; the entire run is visible on three tracks — screen stream + Chain of Thought + step timeline — with prev / next / follow-latest navigation.
Experiments & Research · Self-Enhancing Dev
PLD (Pi-Loop-Dev) is a development runtime that upgrades "multi-agent collaboration" into a "self-enhancing closed loop": starting from a DESIGN.md, it iterates continuously through a Builder → Tester → Analyzer cybernetic loop, stably converging any OpenAI-compatible model (no Claude / GPT quota needed) to an engineering artifact of "all tests pass + spec aligned." It treats the LLM as an uncontrollable actuator and wraps it with a reliable control system via feedback loops — the core methodology Xianma has crystallized in AI engineering.
Experiments & Research · RPA Automation
The RPA Agent Skills System abstracts browser, mobile, JS reverse-engineering, anti-bot, and workflow-orchestration capabilities into a standardized skill set (Skill + MCP dual-shell) that AI Agents can call directly. It integrates open-source engines such as CloakBrowser / Obscura / BrowserCluster / AgentGo / Lamda, hiding multi-engine fragmentation behind a unified interface, so Agents can treat "automation" as composable, reusable tools even against real-world scenarios like strong anti-bot, encrypted params, and mobile control.
Experiments & Research · LLM Security
A CTF-style LLM security lab built on Kolb’s experiential learning cycle — hands-on practice against OWASP LLM Top 10, covering real exploits, flags and defenses for prompt leakage, direct / indirect prompt injection, and SSRF via model output. Each module follows the four-stage pedagogy EXPLORE (blind test) → LEARN (theory) → PROVE (guided exploit) → REFLECT (hardening), runs fully offline via Ollama, and lets security people sharpen AI-security and LLM-pentest skills in a safe environment.