SpecuLab-MCP

• 

  10_printable_optical_card

• 

  09_cool_verification_logs

• 

  01_hero_wow_moment

• 

  08_hardware_simd_benchmark

• 

  04_two_pass_fusion

• 

  06_clean_foil_rejection

• 

  07_evaluation_metrics

• 

  03_supervisor_cloudwatch_gate

• 

  05_unresolved_escalation

• 

  02_agentic_timeline_trace

Inspiration

Metallic foil (blister packs, battery pouch seals) acts like a mirror under directed light. A saturated glare band can hide a defect from a camera. A single-view detector has two options: inspect the glare as-is, or mask it out and accept a blind spot. A human inspector tilts the part. We wanted to automate that reflex.

What it does

SpecuLab-MCP is a prototype active-perception inspector for flat foil:

1. Detects an ArUco marker to rectify each view onto a metric canvas (2000 x 800 px).
2. Measures glare with a two-stage hysteresis mask (core V >= 240, S <= 35, plus its surrounding halo).
3. Exposes three stateless MCP tools: rank_candidate_views, fuse_candidate_view, finalize_inspection_or_escalate. Fusions are capped at 2.
4. Registers an alternate view with homography (USAC_MAGSAC when available), gain-matches it, and feather-blends it over the glare.
5. If glare stays above the gate (2.0%), the case is escalated to a human supervisor instead of being guessed.

Results (read this first)

All results below come from a procedurally generated test set of 14 synthetic specimens (7 defective, 7 clean). No physical hardware or photographs were used. They show the pipeline works on this synthetic set, not that it works on real foil.

• Passive unmasked: strict recall 1/7, clean false rejects 3/3.
• Passive glare-masked: strict recall 0/7, clean false rejects 0/3.
• Active (deterministic): strict recall 5/7, recall including escalation 6/7, clean false rejects 0/3. Small sample; counts shown instead of percentages.

How we built it

• OpenCV for ArUco rectification, morphology, ORB, and homography.
• FastMCP server for the three tools.
• Amazon Bedrock Converse with a deterministic fallback .
• Streamlit dashboard with a before/after comparison, the agent tool-execution timeline, and a benchmark tab.

Challenges

• Blending a clean view into a glare zone left halo artifacts. Fixed by dilating the blend mask before feathering.
• Glare rims were flagged as defects. Fixed with hysteresis masking.

Limitations

• Synthetic data only; no physical validation.
• Persistent glare across all views is escalated, not resolved.
• Small test set.

What we learned

Processing cannot recover information that saturation destroyed. Changing the viewpoint can.

Built With

• active-vision
• aws-bedrock
• boto3
• computer-vision
• fastmcp
• mcp
• metrology
• opencv
• python
• streamlit