AI Lab, running on the same backend as ABI
Take ABI apart. Poke at the pieces.
Every question to ABI travels the same path. Watch it below, then try to break the guard, see what it reads, and match it against a real job.
Try to trick ABI.
Every message is screened by a small classifier (Llama Prompt Guard 2) before the agent ever sees it. Write an attack, or pick one.
The guard's verdict and risk score will show up here.
See what ABI reads before it answers.
Hybrid retrieval over my real data: meaning (pgvector embeddings), exact words (full-text search) and titles, merged with reciprocal rank fusion.
The passages ABI would read, ranked by how well they match.
Match a job description.
Paste a role. ABI pulls the evidence from my real projects and experience, scores the fit and names the gaps, with sources.
what ABI does with it
- Read the requirements
- Find evidence in my work
- Score each requirement
- Be honest about gaps
The answer opens in ABI's panel, with sources.
Under the hood
Free-tier models, real engineering.
- gpt-oss-120b on GroqAnswers and tool calls
- LangGraphThe agent loop
- pgvector on NeonHybrid search over my data
- Gemini embeddingsMeaning as numbers
- Llama Prompt Guard 2Screens every message
- Whisper + OrpheusEars and voice