Closing the books with agents: the exception benchmark
How owned, fine-tuned models behind an independent verifier beat raw-API baselines on straight-through processing and accuracy — with the methodology to reproduce it.
Read the article →- Home
- Blog
Blog
Field notes on autonomous finance
Benchmarks, playbooks, and engineering deep dives from the team building the agentic close.
Categories
BenchmarksEngineeringPlaybooksProductIndustryVoiceSecurityResearchBenchmarksEngineeringPlaybooksProductIndustryVoiceSecurityResearch
Why we publish
Trust is earned with evidence. We share benchmarks and methods so finance and engineering leaders can verify our claims, not just read them.
Latest
Recent writing
The exception benchmark, explained
Our methodology for measuring STP and verifier catch-rate.
Read →Building an independent verifier
Why the model that checks must be separate from the model that acts.
Read →Introducing agents without losing control
A controller's framework for shadow mode to autonomy.
Read →Inside the Closora suite
How five owned products compose one autonomous close function.
Read →Holding a call under 300ms
The realtime pipeline behind Vocora collections.
Read →The data flywheel as a moat
How customer corrections compound into owned-model advantage.
Read →Engineering
From the people building it
Fine-tuning on NeMo
How we train the reasoning and verifier models on DGX.
Serving under 100ms
TensorRT-LLM and Triton in the verification hot path.
Training in simulation
Why Twinn validates every agent before production.
Playbooks
For finance leaders
The 60-day pilot
How to prove agentic close on your own books.
Land and expand
Sequencing workflows to grow net retention.
Auditor alignment
Turning an immutable trail into faster sign-off.
Subscribe
Get new posts
Reading by topic
Find your thread
0+
Articles
0
Benchmarks
0
Playbooks
0
Eng deep dives
Ready when you are
Read it, then see it.
Book a demo and watch the benchmark run on your data.