All projects

Case study 01 / Multimodal AI · Industry internship

Truuth Document Forensics Pipeline

Finding the edits a vision model missed.

All 40 modified documents detected in the final evaluation, with zero false positives.

  • Python
  • AWS Bedrock
  • Qwen3-VL-235B
  • RAG
  • Evaluation

01 / Context

Problem

Document tampering can be visually subtle. A small modified region can disappear inside an otherwise plausible page, making document-level multimodal inference an unreliable signal on its own.

The pipeline needed to combine complementary evidence and track both missed modifications and false alarms as the approach evolved.

02 / Ownership

My Role

As an AI Engineer Intern at Truuth, I built a Python document-forensics pipeline using Error Level Analysis heatmaps, PDF structural analysis and Qwen3-VL-235B through the AWS Bedrock Converse API, with a RAG component.

I automated benchmarking across 12 versions, tracked false positives and false negatives as regression tests, investigated model failure and wrote R&D handover documentation. The internship concluded in June 2026.

03 / System design

Architecture

  1. 01Document input

    Document images and PDFs

  2. 02Forensic signals

    ELA heatmaps + PDF structural analysis

  3. 03Multimodal reasoning

    Qwen3-VL-235B · Bedrock Converse · RAG component

  4. 04Decision & evaluation

    Decision logic + versioned regression checks

Simplified conceptual architecture based on the supplied project description. All displayed documents and signals are synthetic; no employer data or proprietary implementation is reproduced.

04 / Engineering decisions

What Went Wrong & How I Fixed It

The signal was getting lost before the prompt could help.

The vision encoder’s patch averaging diluted fine-grained tampering signals. Small edits were being blended into the surrounding page representation.

I changed the inputs and the decision logic rather than relying only on prompt changes. This addressed the information reaching the model and how the pipeline used the evidence.

Automated evaluation across 12 versions made regressions visible. The debugging lesson: inspect the representation and the surrounding system when a model misses a small but important signal.

05 / Evidence

Results

40/40modified documents detected
0false positives in final evaluation
12versions benchmarked

These are the supplied final evaluation results, not a claim of universal accuracy. The number of genuine control documents and the full evaluation protocol have not been supplied. Specificity and overall accuracy are therefore not calculated.

06 / Tools & capabilities

Stack

  • Python
  • AWS Bedrock
  • Qwen3-VL-235B
  • RAG
  • Evaluation
Next case study

Stix Eatery Platform