Knovyra | Yun

Problem Background

System Flow

One full request lifecycle: user question → Knowledge Flywheel similarity check → Agent Planner → tool-call loop (6 tools, capped at 8 steps) → submit_conclusion → ApprovalRequest (PENDING) → Manager decision at /approvals → write only after approval → AuditLog

Tech Stack

4 layers: Frontend (Next.js 16 + Agent Console + Dashboards) / Application (API Routes + role governance + rate limiting) / Intelligence (Agent Planner + 6 function-calling tools) / Infrastructure (Vercel + Postgres/Neon + Vercel Blob + OpenAI/Anthropic, swappable)

Frontend

Next.js 16 App Router TypeScript / React 19 Tailwind CSS v4 react-markdown + remark-gfm

Application

Next.js API Routes Role Governance Postgres Rate Limiting

AI / Intelligence

Vercel AI SDK (streamText) 6 Function-Calling Tools generateObject Multimodal OCR

Infrastructure

Vercel Serverless Postgres (Neon) + Prisma Vercel Blob OpenAI / Anthropic

Engineering Trade-offs

Cost

Speed

Accuracy

Stability

My Role

Solo Full-Stack & AI Engineer

Built entirely alone, end to end: the data model (Prisma schema across 5 groups), a reproducible synthetic-data generator (fixed RNG seed, with 5 deliberately planted anomaly stories), the agent tool-call loop and its 6 tools, the document-extraction and rule-based cross-validation pipeline, the Knowledge Flywheel (similarity retrieval + reuse tracking), the governance layer (role permissions, approval queue, audit log), 9+ frontend pages and the UI/UX, deployment and CI, and all of the documentation (README, architecture-decision notes). No collaborators, no outside design files.

Features

Operations Agent

Document Intelligence

Knowledge Flywheel

Jira-Style Issue Tracking

Governance & Audit

Challenges & Solutions

1 OpenAI's strict structured-output mode rejects .optional() fields

Problem

Solution

2 A Prisma unique constraint blocked a second approval request

Problem

Solution

3 The "plan" kept reappearing in the stream

Problem

Solution

4 Long JSON strings blew out the tool-call cards

Problem

Solution

5 Unrealistic synthetic-data growth buried a planted anomaly

Problem

Solution

6 The ts-node seed script couldn't resolve @/ path aliases

Problem

Solution

7 A private repo link ended up on the public homepage

Problem

Solution

Expected Impact & System Value

This is a portfolio demo with no real production traffic, so I won't quote "N hours saved" or "N% more efficient" numbers with no real users behind them. What's honest to say instead is what this system concretely proves:

All 5 planted anomalies, found with correct evidence

Supplier price hikes, customer churn, inventory buildup, stockouts, chronically late supplier deliveries — the agent surfaces every one of them, citing the actual database numbers and document content that produced the answer, not a lucky guess.

The governance loop runs end to end

Propose → approve → write → audit — all four steps have real, queryable data behind them. It's not a UI mockup of a flow that doesn't actually happen.

Knowledge reuse verified in practice

Resolve an issue, write down the fix — ask a similar question later, and the agent actually cites that resolution and declines to suggest a duplicate. That's the Jira-style historical-case integration working at small scale, not a claim on a slide.

Document intelligence holds up on messy input

69 synthetic documents mix clean tables, forwarded-email-style casual invoices, handwritten-style receiving slips, and scanned images/screenshots — run through real vision OCR plus structured extraction, not just the cleanest format.

A working answer to a real integration question

The combination of an MCP-shaped tool layer, enforced citations, and an approval gate is a direct answer to "how do you bring agentic AI into an existing system like Jira without losing control" — answered with a running system, not a slide.