The Agent Registry
Every agent here exists because an audit demanded it.
None of these started as "let's build an AI agent." Each one started as a finding, a churn audit, a pipeline audit, a desk drowning in repetitive classification, and became a system. Below: why each agent exists, how it thinks, and how it was hardened against real users.
Select an agent
Six systems. One method.
Four run live at IndiaFilings on real customers. Two are systems built around the same discipline: discovery before recommendation, validation before write, and a named rule for every observed failure.
How they're built
The discipline is the product.
The prompts differ; the engineering method behind all six is identical. Three habits repeat everywhere:
Audit the failure, name the rule
Every bug found in live use, memory loss after a tool call, a re-asked question, a wrong-year document accepted, becomes an explicit, named rule in the next version. Nothing is fixed by vibes.
Validate before you write
Checksums on identifiers, ±₹1 arithmetic reconciliation, tax type derived from law rather than trusted from the document, and nothing posts to a system of record without human confirmation.
Eval, don't hope
The flagship runs a 50-test eval suite plus an LLM-as-judge pipeline scoring six dimensions on every live conversation. Versions ship when the evals say so, not when the demo feels good.
Want the full story?
The details live in a conversation.
Iteration histories, eval design, the knowledge-base architectures, I'll walk you through any of it, including the failures that shaped the rules.
Summarized by design. The full instruction sets, knowledge bases and pricing systems are commercial IP built at IndiaFilings, described here as architecture, never reproduced.