Team Info:
Team 8 - GetThruLaw
Inspiration
Regulatory changes are published constantly, but legal teams still need to manually compare them against policies, checklists and workflows. We wanted to make it easier to identify which internal documents are affected and what needs attention first.
What it does
GetThruLaw allows lawyers to:
- Upload internal documents.
- Automatically categorise and embed them.
- Retrieve recent Singapore Gazette instruments.
- Rank Acts by semantic relevance.
- Compare relevant Acts against document clauses.
- Generate clause-level issues, priority scores, confidence levels and recommended actions.
- Review affected documents through a central Dashboard.
How we built it
We built a browser-based prototype with a local Python server and OpenRouter. The system uses:
- Document text extraction from Word files.
- Qwen embeddings for Acts and document chunks.
- Cosine similarity for relevance scoring.
- Specific legal-term boosting and generic-term penalties.
- DeepSeek for structured impact analysis.
- Local storage for document, Gazette and analysis results.
- A clause-level review interface for lawyers.
Challenges we ran into
- Avoiding falsely high scores from generic legal language such as “Act”, “section” and “regulation”.
- Handling long legal documents within embedding request limits.
- Designing a reliable method to aggregate chunk-level similarity into one Act score.
- Making model outputs consistent enough to generate structured action items.
- Separating document embedding from later legal impact analysis.
- Working with live Gazette data and inconsistent source formats.
Accomplishments that we’re proud of
- Built an end-to-end legal change monitoring workflow.
- Implemented Act-to-document semantic search.
- Added top-three document matches for each Act chunk.
- Created a transparent relevance score rather than relying only on model judgement.
- Generated precise clause-level action items.
- Added priority and confidence scoring.
- Created a lawyer-focused Dashboard for reviewing affected documents.
- Added live Singapore e-Gazette retrieval.
What we learned
Embeddings are useful for finding conceptual matches, but cosine similarity alone is not enough. Generic legal language can create misleading scores, so relevance needs additional term weighting, filtering and calibration. We also learned that AI should identify and explain potential issues, while deterministic code should handle scoring, thresholds and workflow decisions wherever possible.
What’s next for GetThruLaw
- Improve chunking by splitting legislation by section and subsection.
- Add strict JSON Schema outputs for more reliable model responses.
- Support model selection for speed, cost and accuracy.
- Add better source citation and side-by-side legal evidence.
- Improve live Gazette filtering and monitoring.
- Add review history, audit trails and document version comparison.
- Introduce human feedback to improve relevance ranking.
- Add secure team workspaces and enterprise data controls.
Log in or sign up for Devpost to join the conversation.