PRISM
Pattern Recognition & Insight Structure Module
AI-560: AI Design & Development Lab, SCAD — 35 days, 12 development sessions
Role: Solo Developer
Tools: Python, HP AI Studio, NVIDIA RTX PRO 6000, Gradio, Whisper, Figma API, vector embeddings / semantic clustering
The Problem:
Research teams spend 3-4 weeks manually synthesizing data across 7 steps and 5+ platforms. Sprint after sprint gets delayed because research can't keep pace with development velocity. I experienced this firsthand and couldn't find a single tool that solved it.
1
The Solution
A multi-modal pipeline that ingests FigJam boards (via Figma API), audio transcripts (via Whisper), and PDF/PPT documents — then runs them through a custom pattern recognition engine that clusters and synthesizes findings, with every insight traceable back to its source.
2
System Architecture
Synthetic Data Generation
Built a custom generator (~800 lines) producing 899,195 labeled sticky notes across 10,000 boards, with 6 researcher personas modeling realistic behavioral variation — typo rates, color consistency, spatial clustering — across 5 industries. Originally planned to scrape real FigJam boards; pivoted to synthetic data after identifying that scraping would violate Figma's ToS and raise academic integrity concerns. The pivot gave full label control and let edge cases (15% atypical color usage) be built in deliberately.
Pattern Recognition Engine
PRISMBrain (~1,200 lines) uses semantic analysis and vector embeddings to read sticky note content directly, rather than relying on color-coding conventions. Early assumption was that color maps to meaning (yellow = neutral, red = pain point) — real researchers don't follow that consistently, so the system reads content semantically and treats color as supplementary context only.
Generator-Learner Feedback Loop
3
Two AI systems train each other: the Generator creates synthetic boards, the Learner processes them, and confidence scores feed back to produce harder or more varied examples — allowing the system to improve without retraining from scratch.
Source-Attributed Output
Every insight links back to where it came from, with human validation built in as a requirement, not an afterthought. This was non-negotiable from day one — if PRISM couldn't show a researcher where an insight came from and let them reject it, it would reproduce the same opacity as existing tools (Dovetail, Miro AI, Notion AI, Figma AI).
Outcome
A working production system deployable via Gradio share links. End-to-end demo: a FigJam board with 688 notes processed with source URL captured, a PDF analyzed into 12 categorized and prioritized insights, and a full glassmorphic interface with sidebar navigation and tabbed data sources. Pattern recognition accuracy: 90%+ on the synthetic test set. Synthesis time: 11+ hours down to approximately 2.
This was an academic capstone with no public deployment or external user cohort at submission — the project demonstrates the full arc from ethical foundation through production interface, including the judgment to stop and redesign when the original approach raised a legal and ethical problem.
5
Next Steps
Next steps for PRISM involve further User testing with an authentic data pool and future uploading to Huggingface
5