Make privacy non-skippable
The controller runs PIIAgent before the model stages so user metadata is masked before AI processing.
AI / ML · Academic group research · Case study
Privacy before
Privacy before
prediction.
A four-agent research pipeline for social media credibility assessment. My contribution was the privacy layer, orchestration, routing, audit records and API integration; teammates developed the classification and similarity models.
ACACAS — Automated Credibility Assessment System cover
4 agentsPII → text → image → fusion
3 input modesText, image or both
Group researchMy contribution clearly identified
How a post travels through ACAS
Privacy is handled before anything else sees the content.
1
Post submitted
A claim arrives through the API
2
Privacy masking
Personal details are removed first
3
Controller routing
The request goes to the right agents
4
Credibility and similarity
Specialist agents assess the claim
5
Audit record
Each decision is stored for review
The problem
Credibility checks can expose the people being checked.
Social media posts contain both content to assess and personal metadata that should not be sent through every model. ACAS was designed to route text-only, image-only and multimodal posts through a sequential assessment while masking user metadata before analysis. This was a group final-year research project at SLIIT.
Architecture
Privacy is the first stage, every time.
The diagram shows how data moves through the project.
Post inputText, image, or both
PIIAgentMandatory masking
Model agentsRoBERTa + CNN
FusionCLIP + OCR
VerdictExplanation + audit
The diagram summarizes a group-built pipeline. I designed the PIIAgent, controller, routing, audit architecture and /verify integration. Teammates developed the text, image and multimodal models.
Key decisions
Why I built it this way.
Route by available input
Text-only, image-only and multimodal posts follow applicable stages through a shared orchestration interface.
Keep raw and masked evidence distinct
MongoDB collections tweets_raw and
pii_audit record the input and privacy
transformation for review.
Results
What was delivered.
Scope and measurement matter as much as the headline numbers.
Orchestrated pipeline
CrewAI controller enforcing agent order and plugging team-built models into a common flow.
Privacy layer
PIIAgent masks username, name, location and profile-image metadata before analysis.
API and records
FastAPI /verify endpoint, Pydantic schemas and
MongoDB audit collections.
Measurement: This page focuses on the architecture and privacy controls. It does not report model accuracy, precision, recall or latency.
What broke
What I learned while building it.
Integration boundaries
Four independently developed components needed a shared controller and plug-and-play interface. The lesson was to agree one shared interface early, so each component can be swapped without rewriting the controller.
Audit visibility
Raw and masked records were separated to make the privacy transformation inspectable. Access controls for those records are not documented.
Model delivery
Large model weight files are excluded from the public repository; local reproduction requires the weights or retraining instructions.
Limitations
What this project does not claim.
This is an academic group project. The repository
omits large model weights and does not publish benchmark metrics yet. Docker deployment, live social-media ingestion,
multilingual support and a browser extension are listed as future
work. Masked PII and raw audit records require careful access
control in any real deployment.
Explore the work
Code, report and next steps.
The repository contains the orchestration code and project setup. The work was developed for a conference-paper submission; no publication acceptance is claimed.