Local-first enterprise knowledge agents, delivered FDE-style.

Pain points

• Rules and regulations scattered across the institution — compliance questions meant digging through files.
• Reviews done by hand: slow, and inconsistent between reviewers.
• SOE-grade requirements: private deployment, Xinchuang stack, MLPS 2.0 Level 3.

Solution

We structured the entire rules system into a clause-level knowledge base, then built clause-level traceable compliance Q&A, document review, writing assistance and regulation comparison on top — FTS5 + vector hybrid recall with evidence grading and refuse-without-basis, a primary/fallback dual-model link that runs fully on the intranet, and a SHA-256 audit chain. Four business agents were developed on site: 7-stage procurement, procurement co-pilot, deep contract review and an expert panel.

Pain points

• Three enterprise accounts (2 Xiaohongshu + 1 Douyin) — trend-watching and drafting all by hand.
• Accounts can't be logged into; only external public-data collection is allowed.
• Daily output must stay stable, in each account's own voice.

Solution

Compliant collection via real-browser automation (no cookies read, no captcha bypass) builds an editable tree-structured knowledge base; a cross-platform hot-topic board clusters trends into use-now / watch / caution; Kimi k3 produces daily scripts and posts in a structured pipeline; a multi-agent review stage only passes pieces scoring ≥85; interactive revision, plus a daily brief in the tenant's timezone.

Pain points

• The 78-page scanned Philippine national standard PNS 25:1994 had to be structured fully automatically.
• Every parameter must carry clause number, verbatim English evidence, page number and confidence.
• Nothing may be fabricated.

Solution

PyMuPDF 160-DPI rendering + RapidOCR local recognition, with tables reconstructed from coordinates; a DeepSeek reasoning model extracts clause by clause over the full text; deterministic checks plus LLM re-review run multi-round quality control with re-extraction on failure; anything inconsistent with human values is flagged “needs human review”.

Pain points

• Finding creators for promotion meant manually browsing platforms.
• No system between a one-sentence need (“30 couple bloggers on Xiaohongshu, 50k+ followers, ¥20k budget”) and a contactable list.

Solution

Natural-language needs → structured search criteria (rule-based, with Kimi AI parsing) → compliant collection of public profiles via WebBridge → S–D engagement grading with CPM estimates → filter, follow up, and export business materials. A zero-dependency local app, double-click to run.