AI Lab
Experiments between projects.
Some ended up in client work, and some taught us what not to do. The numbers come from our own benchmarks, and each page explains how we measured them.
10 experiments
- LLMLive
85% bug detection rate
Local LLM code reviewer
Code review from models running on your own machine. No code is sent to an external API.
Qwen 2.5 72B, Llama 3.3 70B, CodeLlama
Read the write-up - AgentsLive
94% query accuracy
Voice-to-SQL agent
Ask a database a question out loud and get back the query, the result and a short explanation.
GPT-4o, Whisper Large-v3
Read the write-up - RAGLive
12 metrics
RAG evaluation harness
A benchmark suite that compares chunking strategies, embedding models and retrieval methods on 12 metrics.
Claude Sonnet, GPT-4o, all-MiniLM-L6
Read the write-up - AgentsLive
3 rounds per debate
Multi-agent debate
Three agents argue different sides of a question. A judge agent then weighs the arguments and writes a conclusion.
Claude Sonnet, GPT-4o, Gemini 2.5 Pro
Read the write-up - VisionLive
96% field accuracy
Document vision parser
Pulls structured fields out of receipts, invoices and forms with vision models, without OCR templates.
GPT-4o Vision, Claude Vision
Read the write-up - AgentsResearch
3 memory types
Agent memory
A long-term memory layer for agents, with episodic, semantic and procedural memory and a forgetting curve.
Claude Sonnet, all-MiniLM-L6
Read the write-up - LLMLive
94% accuracy
Real-time sentiment stream
Sentiment scoring on streaming text such as social feeds, chat and support tickets, in under 100 ms per message.
DistilBERT (fine-tuned), RoBERTa
Read the write-up - RAGLive
Entity extraction F1 of 0.89
PDF to knowledge graph
Turns a PDF into a browsable graph of entities, relationships and the passages they came from.
Claude Sonnet, GPT-4o
Read the write-up - AutomationLive
78% task completion
Browser automation agent
Give it a goal in plain English and it drives a real browser to navigate sites, fill in forms and collect data.
GPT-4o Vision, Claude Sonnet
Read the write-up - LLMResearch
15% average accuracy gain
Prompt optimization lab
Improves prompts automatically with a genetic algorithm, then checks each gain with an A/B test against an eval suite.
Claude Sonnet, GPT-4o, Gemini 2.5 Pro
Read the write-up