AccordWorks
Smarter way to review contracts
Created on 29th August 2025
•
AccordWorks
Smarter way to review contracts
The problem AccordWorks solves
The Problem It Solves
Contracts are everywhere: employment agreements, vendor contracts, NDAs, SaaS subscriptions, freelance work, and even rental agreements. But reading and understanding them is often time-consuming. Pages of dense legal jargon slow down decisions.
Confusing: Hidden clauses, tricky wording, and fine print make it easy to miss risks.
Expensive: Hiring a lawyer for every contract review is not always practical.
Risky: Overlooking a single clause can lead to financial or legal trouble later.
How AccordWorks Helps
AccordWorks makes contract review simple, fast, and safe by using AI to:
- Summarize contracts into plain language you actually understand.
- Highlight key clauses (termination, payment, liability, confidentiality, etc.).
- Flag potential risks or uncommon terms you should watch out for.
- Provide insights & suggestions so you can make better, faster decisions.
- Save hours of manual reading and avoid costly oversights.
Who Can Use It
- Freelancers & Creators—Quickly review client contracts before signing.
- Startups & Small Businesses – Check vendor and partnership agreements without hiring a lawyer every time.
- Job Seekers—Understand your offer letter or employment contract.
- Landlords & Tenants—Simplify rental agreements.
Anyone dealing with confusing legal documents who wants clarity and confidence.
Challenges I ran into
Unreliable document parsing (PDF/DOCX diversity)
Some PDFs were image-only or had broken structure; DOCX styles varied wildly.
Fix: Added a parsing pipeline with fallbacks (native text → layout-based → OCR for scans), preflight validation, and normalized outputs to a common schema before downstream AI steps.
LLM hallucinations in clause extraction
Early extractions invented clauses or mislabeled fields.
Fix: Switched to schema-constrained generation (tool/function calling + JSON schema), added validators and auto-retry with minimal few-shot exemplars, and flagged low-confidence outputs for re-prompt.
Long contracts vs context limits
Full documents blew past model context and degraded accuracy.
Fix: Semantic chunking with overlap, hybrid retrieval (BM25 + embeddings), and map-reduce for summaries. Included per-chunk citations so users can trace outputs.
Risk scoring calibration
Scores were inconsistent across contract types due to sparse labels.
Fix: Built weak-supervision rules for bootstrapping, then calibrated with a small vetted set (e.g., Platt scaling). Exposed confidence bands and rationales per risk.
Latency and rate limits
Parsing + RAG + multiple model calls caused slow UX and occasional throttling.
Fix: Async background jobs with streaming UI updates, response caching (parses/embeddings), batched calls with exponential backoff, and idempotency keys.
Privacy and compliance handling
Logs accidentally captured snippets; storage needed hardening.
Fix: PII scrubbing in logs, encrypted storage, short-lived presigned URLs, least-privilege access, and an optional “local processing” mode for sensitive docs.
Tracks Applied (1)
All Registered Participants
Windsurf
Technologies used