Project presentation
Notes from the original project deck, kept with the decisions and trade-offs.
Orchestrating a digital newsroom of expert agents to proofread at scale.
This wasn't a startup idea; it was a tuition fee.
I built this to understand the "Hidden Taxes" of agentic workflows: latency, token costs,
and orchestration fragility.
Getting three LLMs to agree on a diff is
harder than managing three human interns. Learning CrewAI revealed how quickly
context windows fill up when agents "talk" to each other.
Elevator Pitch: Defining "Personas" to constrain LLM behavior. Solution: Implementing a hierarchical process
where agents submit proposals, but only the Manager agent has write access to the final
document. Reduction in "first pass" editing time. Taught me the value of frugal prompting. Deep dive into CrewAI architecture. (CLI Output Visualization) Agent orchestration & task delegation. High reasoning capability for nuance. Text extraction and chunking.01 CREWAI EDITOR+
CREWAI
EDITOR
02 The Cost of Complexity+
The Cost of
Complexity
03 Digital Newsroom+
Digital Newsroom
A CLI-based multi-agent system where specialized AI
personas
(Grammarian, Stylist, Fact-Checker) critique documents in parallel.
Approach
Pros
Cons
Verdict
Single LLM (ChatGPT)
Cheap & Fast.
Lacks nuance; prioritizes speed over rigor.
Baseline
Human Editor
High Agency.
Expensive ($50/hr); Slow turnover.
Too Slow
Agent Crew
Multi-perspective critique.
High Token Cost.
Selected
04 CrewAI Framework+
CrewAI
Framework
editor = Agent(
role='Senior Editor',
goal='Fix tone inconsistencies',
backstory="""You are a veteran editor at wired.com.
You prefer punchy, active sentences.""",
verbose=True,
allow_delegation=True
)
task = Task(
description=f"Critique this section: {text_chunk}",
agent=editor
)05 The Hallucination Loop+
The Hallucination
Loop
06 The "Lesson"+
The "Lesson"
~80%
HIGH
Learned
07 Terminal Demo+
Terminal Demo
08 Stack Details+
Stack Details
CrewAI
GPT-4o
PyPDF2
09 STATUS: ARCHIVED+