About
Named for the stone that does the judging.
A small, senior, evidence-first practice. Every claim traceable to a trace.
βάσανος · basanos
/bá.sa.nos/ · noun, Ancient Greek
1. The touchstone — the dark stone assayers dragged gold across to read the streak and judge purity.
2. By extension: the test itself. The trial that reveals what a thing actually is.
Who we serve
PE portfolio companies shipping agents into regulated, unforgiving workflows.
Loan servicing, claims, collections, compliance-heavy operations — places where a wrong answer isn’t a bad demo, it’s exposure. Operators there don’t need another dashboard; they need a straight answer on whether the agent is safe to ship, and a gate that enforces it.
How we work
Small, senior, evidence-first.
Every engagement starts at the stone: a two-week eval audit with scored traces you can read yourself. Every claim we make is traceable to a trace. And every deliverable is built to feed the next altitude — the harness produces the data, the data hardens the judge, the judge trains the model you end up owning.
Sariph Shrestha · Founder · Washington, DC
The assayer
A decade of systems where wrong answers cost money.
Basanos is run by Sariph Shrestha, an AI engineer whose career has stayed inside regulated, unforgiving workflows: fraud detection at Capital One (34% fewer false positives at 99.7% detection), a HIPAA-compliant healthcare platform serving 3.2 million members at CareFirst, loan-processing systems at Fannie Mae.
Most recently, the retrieval and reasoning stack for enterprise LLM products at IBM — hallucination mitigation with citation enforcement, faithfulness evaluation, regression testing that catches drift across model updates, multi-agent orchestration with human-in-the-loop checkpoints, and model serving at 10M+ daily inferences. Basanos is that discipline, offered as a practice.
- Focus
- LLM evals · judge & reward hardening · SFT and RL post-training
- Track record
- Capital One · CareFirst BlueCross BlueShield · Fannie Mae · IBM — a decade of regulated, high-stakes systems
- Education
- B.S., Electrical & Computer Engineering — The University of Texas at Austin
Disclosure
Meridian Servicing and its agent Harper, used across this site, are composite illustrations — details drawn from multiple engagements and published practice. The numbers are plausible, not measured; the failure modes and methods are the real ones.
We label it because the thing we sell is honest judgment under pressure — “the judge that survives its own audit” applies to our marketing too.
Drag your agent across the stone.
A two-week eval audit. Scored traces, a regression harness, and a straight answer.
hello@basanos.ai · Replies within 48h