0
AVERNUS PROJECT

Fello

AI-Powered Project Planning & Management Workspace.

Reduced initial project setup time from hours of manual entry to seconds via live streaming parsers.

Client / ScopeInternal R&D / SaaS Client Product
Timeline4 Months Build
RoleLead Full-Stack Architect & Streaming UI Designer
CategoryAVERNUS PROJECT
[ HERO IMAGE: Fello Interface ]

Replace with high-res UI dashboard screenshot, mockup, or banner (Recommended 1920x1080)

Production Ready · Avernus Core Project

Production-Grade AI Workspace

Fello bridges natural language planning and structured agile execution. Users describe project goals conversationally, and Fello's custom streaming buffer parser intercepts JSON payloads in real time to generate interactive Kanban boards with on-demand AI task companions.

Eliminated raw JSON UI leaks through custom zero-dependency streaming buffer parsers
Interactive Kanban board with independent column scrolling and @dnd-kit drag-and-drop
AI Task Drawer generating technical specs, QA test plans, and implementation steps per card
03 · Executive Overview

Project at a Glance

A premium AI workspace that translates natural language project descriptions into structured Scrum boards, equipped with an AI Project Companion drawer for engineering specs, test plans, and technical steps.

Domain & Focus

Agile Project Planning & Generative Management

Stakeholders

Modern Engineering & Product Teams

Engagement

Q2 - Q4

Primary Win

Instant translation of ambiguous requirements into structured epics, features, and Jira-style task cards.

04 · Operational Bottleneck

The Problem

Product managers and engineering leads spend countless hours drafting epics, defining features, and creating task cards. Generative AI tools output raw text or malformed JSON that leaks into user interfaces, leaving tasks disconnected from real execution environments.

Friction Point #1

Generative text LLMs leak raw JSON syntax into streaming chat bubbles, disrupting UX.

Friction Point #2

Task boards created by AI are typically static and cannot be dragged, reorganized, or persisted.

Friction Point #3

Developers lack detailed technical steps, edge cases, and QA criteria on auto-generated task cards.

Friction Point #4

Fragile client state loses in-progress plans on accidental reloads.

05 · Strategic Targets

Goals

Technical Goals

  • Zero-dependency streaming Markdown bubble renderer compatible with React 19.
  • Buffer parser algorithm to intercept and suppress raw JSON payloads before UI render.
  • Interactive Kanban board with independent column scrolling and dnd-kit integration.
  • Split Task Workspace Drawer generating specifications, QA test plans, and technical blueprints.

Business Goals

  • Accelerate project discovery and backlog initialization by over 80%.
  • Provide engineering teams with actionable, code-ready task specifications.
  • Deliver a fluid, high-performance dark-mode experience with zero layout shift.
06 · Ownership & Execution

My Role

Position

Full-Stack AI Architect

Core Responsibilities

  • Architected the streaming parser buffer to intercept plan payloads in real-time.
  • Built the React 19 Kanban board using @dnd-kit/core and Framer Motion.
  • Engineered the AI Task Companion drawer and prompt chains for specs, test cases, and steps.
  • Designed MongoDB schemas for Projects, Epics, Features, and Tasks.

Primary Focus Areas

Real-Time Stream Buffer ParsingReact 19 Concurrent UI & Drag-and-DropDatabase Modeling in MongoDBPrompt Engineering for Technical Specs
07 · The Architecture Approach

Solution

We engineered Fello with a streaming-first architecture. A custom buffer parser scans live LLM chat streams, cleanly truncates human-readable text before plan definitions start, parses the plan JSON in memory, and immediately renders it into an interactive Kanban board.

PILLAR 01

Stream Buffer Interceptor

Suppresses raw JSON logs from leaking to the UI and converts pending states into interactive animated tool cards.

PILLAR 02

Interactive Scrum Workspace

Drag-and-drop Kanban columns with parent Epic/Feature grouping pills and persistent lock states.

PILLAR 03

AI Project Companion Drawer

On-demand modal generating comprehensive Technical Specs, QA Test Cases, and implementation blueprints per card.

08 · Systems Engineering

Architecture & Data Flow

End-to-end flow from natural language prompt to streaming buffer interception, MongoDB upsert, and interactive Kanban synchronization.

01

User Prompt Ingestion

Next.js App Router / SSE

User describes application requirements in natural language through the AI Planner chat.

02

Streaming Buffer Parser

Custom Buffer Algorithm

Client-side algorithm detects structured JSON boundaries, suppresses raw code, and buffers plan objects.

03

Database Sync & Lock State

MongoDB Driver / Next.js API

Parsed Epics, Features, and Tasks are upserted into MongoDB collections; lock status persisted to localStorage.

04

Interactive Kanban Synchronization

@dnd-kit / Framer Motion

Kanban board dynamically maps cards to columns with drag-and-drop status mutation.

05

AI Companion Drawer Generation

Gemini / OpenAI API

Clicking any card invokes prompt pipelines generating specs, QA checklists, and engineering steps.

09 · Capabilities

Core Features

Zero-Dependency Stream Parser

Zero raw JSON leaks

Intercepts raw JSON payloads from live LLM streams, cleanly rendering conversational text while piping structured plans to state.

Split Task Workspace Drawer

On-demand deep specs

Dual-column editor combining task metadata form with AI-generated Technical Specs, QA Test Cases, and Technical Steps.

High-Density Kanban Board

dnd-kit integration

Independent column scrolling with parent Epic/Feature visual pills and full touch/drag support.

Dynamic Model Switcher & Key Encryption

Multi-provider support

Allows users to switch between Gemini, OpenAI, and Anthropic models with encrypted database key storage.

10 · Tradeoffs & Rationale

Engineering Decisions

Client-Side Buffer Parsing vs Server Parsing

Chosen Path:Client-Side Streaming Interceptor
Alternative Considered:Server-Side Post-Processing

Why: Allows immediate conversational text streaming to the user while asynchronously processing board generation without adding server hop latency.

MongoDB Flexible Document Schema vs Relational

Chosen Path:MongoDB Native Driver
Alternative Considered:PostgreSQL Relational Schema

Why: Epics, features, and dynamic AI-generated technical specs vary in depth and structure; document storage allowed rapid schema evolution.

11 · Obstacles & Solutions

Challenges

Challenge #1

Streaming Truncation Leaks: LLMs streaming incomplete JSON tokens caused UI crashes in standard JSON.parse.

Engineering Solution

Implemented a regex token boundary scanner that truncates the displayed stream precisely before the first JSON bracket and waits for complete payload closure.

System Impact

100% clean UI streams with zero visible code artifacts.

Challenge #2

Kanban Drag Re-renders: Dragging cards in high-density columns caused lag across the entire React tree.

Engineering Solution

Isolated drag state using @dnd-kit sensors and React 19 optimistic state updates.

System Impact

Silky 60fps drag-and-drop interactions across 100+ task cards.

12 · Roadmap & Milestones

Implementation Timeline

Phase 1: Chat & Buffer ParserWeeks 1 - 4

Stream Parser & Markdown Engine

  • Zero-dependency React 19 Markdown renderer
  • Streaming buffer parser algorithm
  • Basic chat interface
Phase 2: Kanban & Data ModelingWeeks 5 - 9

Interactive Board & MongoDB Schema

  • Kanban column layout with @dnd-kit
  • Projects, Epics, Features, and Tasks collections
  • Column drag-and-drop status mutations
Phase 3: AI Companion DrawerWeeks 10 - 13

Task Drawer & Spec Generators

  • Split task workspace modal
  • AI prompt pipelines for specs, tests, and steps
  • Persistent storage of generated specs
Phase 4: Polish & Model ManagementWeeks 14 - 16

Key Encryption & Analytics Dashboard

  • Encrypted API credential storage
  • Workspace KPI analytics dashboard
  • Production release and deployment
13 · Measurable Performance

Results & Metrics

85%
Time Saved

Reduction in initial backlog creation and sprint planning time.

0 leaks
Stream JSON Leaks

Zero raw JSON payload leaks in streaming UI via custom buffer parsers.

<100ms
Card Drag Latency

Instantaneous optimistic status transitions across columns.

3 in 1
AI Specs Generated

Technical Spec, QA Test Checklist, and Implementation Steps per task.

14 · Visual Gallery

Screenshots

[ SCREENSHOT #1: AI Planner Chat & Real-Time Plan Generation ]

Drop high-resolution interface screenshot here (Recommended 1920x1080)

AI Planner Chat & Real-Time Plan Generation

Natural language conversational interface streaming agile sprint definitions.

[ SCREENSHOT #2: Interactive Kanban Board with Drag-and-Drop ]

Drop high-resolution interface screenshot here (Recommended 1920x1080)

Interactive Kanban Board with Drag-and-Drop

Multi-column Scrum board with parent epic tags and status updates.

[ SCREENSHOT #3: AI Task Companion Drawer (Technical Specs & QA) ]

Drop high-resolution interface screenshot here (Recommended 1920x1080)

AI Task Companion Drawer (Technical Specs & QA)

Split drawer generating architectural blueprints and test plans.

15 · Walkthrough

Demo Video

[ DEMO VIDEO PLACEHOLDER ]

Fello Product Walkthrough: 2-minute demonstration showing prompt-to-kanban board generation, card reordering, and the AI Task Spec drawer.

Add iframe or video src in lib/case-studies-data.ts

Demonstration Highlights:

Prompt to full Kanban board in 15 seconds
Drag and drop column status progression
Generating QA test checklists on a single click
16 · Retrospective

Lessons Learned

1

Separating user-facing chat streaming from background structured data parsing provides a vastly superior user experience.

2

Optimistic UI updates are mandatory for drag-and-drop boards to feel native and responsive.

3

Developers value concrete QA test cases and technical steps far more than generic high-level AI advice.

17 · Technologies

Tech Stack

Frontend

Next.js 16 (App Router)React 19Tailwind CSSFramer Motion@dnd-kit/core

Backend & Database

Node.js 20+MongoDB DriverNext.js Route Handlers

AI & Models

Google Gemini APIOpenAI GPT-4oCustom Buffer Stream Parser
18 · Organizational Value

Business Impact

Fello transformed product planning for modern software teams, bridging the gap between high-level vision and actionable engineering execution.

Eliminated hours of manual story writing during sprint kickoffs.
Standardized technical specification quality across diverse feature sets.
Created an intuitive, delightful planning experience praised by product managers.

Fello eliminates the tedious busywork of agile planning. It takes three sentences and turns them into an engineering-ready board with test plans already written.

P
Product Engineering Lead
Avernus Tech Stack
Scale Your Architecture

Have a complex system to build?

Let's engineer your vision. Whether you need multi-agent AI orchestration, full-stack streaming hubs, or hardened enterprise analytics, Avernus delivers scalable architecture from zero to one.