SP Manager
AI-powered project management for engineering teams.
SPManager automates the hardest parts of project delivery — requirement analysis, task assignment, and progress tracking — so teams can focus on building instead of managing.
The Story
Why this exists.
The Problem
Engineering teams spend 20–30% of their time on project management overhead — writing tasks, assigning work, chasing status updates. Most PM tools make this worse, not better.
Why It Matters
Every hour a senior engineer spends in status meetings or updating Jira tickets is an hour not spent building. AI-driven project management is the only way to scale delivery without scaling headcount.
The Solution
SPManager uses RAG to analyse requirements and auto-generate structured tasks, a skill-graph to assign work to the right engineer, and real-time Socket.io sync so everyone sees the same board state.
Product Features
What it does.
AI Requirement Analysis
Paste a requirements document — SPManager extracts tasks, acceptance criteria, dependencies, and effort estimates using a RAG pipeline over your project context.
Skill-Based Task Assignment
A graph of team skills and current workload surfaces the optimal assignee for every task. No more 'who should do this?' in Slack.
Real-Time Board Sync
Socket.io ensures all board changes propagate to every team member within 50ms — no page refresh, no stale state.
Velocity & Burndown
Automatic sprint velocity tracking and burndown chart generation — no manual data entry required.
GitHub Integration
Link PRs and commits to tasks automatically. Task status updates when a linked PR is merged.
System Architecture
How every layer connects.
Requirements → AI extraction → task graph → skill match → assigned task → real-time board update
Technical Deep Dive
Under the hood.
RAG-Based Requirement Analysis
When a user pastes requirements, the system chunks the text, embeds each chunk using text-embedding-004, and retrieves the most relevant prior tasks and project context from the vector store. This context is injected into the Gemini prompt alongside the new requirements. The output is a structured JSON of tasks with title, description, acceptance criteria, estimated effort (S/M/L/XL), and dependency edges.
Skill Graph Assignment
Each team member has a weighted skill vector built from their commit history, past task labels, and self-declared expertise. When a new task is created, its required skills are extracted from the task description. A cosine similarity match between the task skill vector and each team member's profile returns a ranked list of candidates, adjusted for current workload.
Real-Time Consistency
The Socket.io server runs on Railway with a sticky session load balancer. All board mutations go through a command queue — each command is applied optimistically on the client and confirmed or rolled back by the server within one round-trip. Conflict resolution uses vector clocks per task.
// Server: apply command and broadcast
socket.on('task:move', async (cmd) => {
const result = await applyCommand(cmd);
if (result.ok) {
io.to(cmd.projectId).emit('task:moved', result.task);
} else {
socket.emit('task:rollback', { id: cmd.taskId, state: result.current });
}
});Engineering Decisions
Challenge → Decision → Result.
Performance & Scale
By the numbers.
Deployment & Infrastructure
Production setup.
Deployment
Next.js frontend on Vercel Edge CDN. Socket.io server on Railway with sticky sessions. MongoDB Atlas M10 cluster in us-east-1. Cloudflare R2 for file storage.
CI/CD
GitHub Actions — type-check + lint on PR, E2E Playwright tests on merge to main, auto-deploy on green CI.
Monitoring
Vercel Analytics + Web Vitals. Railway metrics for Socket server CPU/memory. MongoDB Atlas Performance Advisor for slow queries.
Challenges & Failures
What broke — and how I fixed it.
What I Learned
Key takeaways.
- 01
RAG quality depends more on chunking strategy than model choice — invest in the chunker first.
- 02
Skill-based assignment only works if skill data is maintained — build the data collection UX before the algorithm.
- 03
Sticky sessions on Railway are not sticky after a dyno restart — need Redis session store for true persistence.
- 04
Real-time features should be designed for eventual consistency from day one.
Future Roadmap
Where this is going.
v2.0 — Q4 2026
- —Fine-grained RBAC — project admin, contributor, viewer roles
- —AI sprint planning — auto-populate sprint from backlog by velocity
- —Slack/Teams integration for task notifications
v3.0 — 2027
- —Multi-tenant SaaS with organisation isolation
- —AI retrospective generation from sprint data
- —Mobile app for async task updates
Screenshots
The product.
Real-time Kanban board with AI-generated tasks
AI requirement → task extraction
Technology Stack
Built with.
Frontend
Real-Time
AI/ML
Database
Storage
Integration
Deployment
Ready to dive in?