Updates & Changelog

We studied dozens of real, high-quality EPBs and completely rewrote how every Air Force format generates bullets. This is the single biggest quality improvement since launch.

• Every format now trained on real performance writing patterns
• Bullets now use the correct 2-sentence structure per MPA block
• Replaced all generic examples with proven, high-quality writing
• Character targets refined to match what actually gets approved (340-350 sweet spot)
• Name variation built in: rank+name, first name, and pronouns across sentences
• Ampersand (&) usage for natural military brevity

Not just EPB/OPB - every single Air Force generation format was improved with real-world writing patterns.

• EPB/OPB Tailored: All 4 MPAs rewritten with proven bullet patterns
• HLRA: Updated to 225-250 character range, stratification now optional
• Duty Description: Char limit corrected to 450, rank-appropriate guidance added
• Traditional: Compression rules rewritten, abbreviation dictionary expanded
• ALQ Focused: Each quality mapped to real performance examples
• Narrative v1 & v2: Both updated with correct 2-sentence structure
• EPB Standard: Real examples for every MPA section

Switched from Google Gemini to Mistral Large for better writing quality and reliability.

• Mistral Large offers excellent instruction following for military writing
• More reliable than Gemini with no quota issues
• Select from: Groq (fast & free) | Mistral (quality) | OpenAI (premium)
• Existing Gemini users automatically migrated to Groq

The workbench expand and contract buttons now use AI to intelligently adjust your bullet length while preserving quality.

• Expand: AI adds metrics, scope details, and impact language to hit character minimums
• Contract: AI removes filler, applies abbreviations, and tightens language
• Both maintain military voice and MPA compliance
• Streaming responses so you see results in real-time

Hitting regenerate now gives you genuinely different bullets every time. The system bans previously used opening verbs and their entire synonym families.

• Opening verbs from previous bullets are now banned (with synonyms)
• 2nd attempt: completely different structure and emphasis
• 3rd+ attempt: bold, unconventional angles on the same accomplishment
• Temperature dynamically adjusts per variation type for optimal creativity

Added in-memory rate limiting to protect the service and ensure fair usage for everyone.

• 30 requests/minute for Generate and Improve
• 15 requests/minute for Award generation
• Per-IP tracking with automatic cleanup
• Clear error messages with retry timing when limits are hit

You can now see exactly which AI model is generating your bullets. No more silent fallbacks.

• Response headers show which provider actually ran your request
• If your selected provider isn't available, you'll know it fell back to Groq
• Activity logs track which provider was used for every generation

Hardened the application with better secret management and input handling.

• Moved sensitive configuration to environment variables
• Improved error handling across all API routes
• Better quota and rate limit error messages with actionable suggestions