Writing by Kaushal Khodifad, founder of CLOZOM: long-form teardowns of how AI systems actually get shipped inside regulated industries, covering compliance, audit trails, latency budgets and real users. 5 teardowns are published, 97 minutes of reading across 85 cited sources. Every factual claim carries a citation to a primary source.

Teardowns, not takes.

Long-form on how AI actually gets shipped inside regulated industries: compliance, audit trails, latency budgets. Every claim carries a source you can open.

5 teardowns97 minutes of reading85 cited sources

  1. Grounding is the easy half of autonomous medical coding - an Arintra teardown

    Validating every code against the actual CMS tables stops a model inventing codes, and that fight is largely won. It does not stop a real row from being the wrong answer, which is the failure mode nobody demos.

    Healthcare revenue cycle · Medical coding · Retrieval grounding · CMS code sets · Denials

    Working version: Arintra

    Read the teardown

    20 min read

  2. How card-data entitlements actually work - an Atlas teardown

    An entitlement that lives in application code is a claim about every code path that will ever exist. An entitlement that lives in the database is an artifact somebody else can read.

    Payments data · Row-level security · Postgres · Audit evidence · Governance

    Working version: Atlas

    19 min read

  3. Compliance evidence: automating the part auditors actually check - DPDP Rules 2025 teardown

    A regulator has already said it plainly: pointing at a correctly configured system is not proof that it was configured correctly on the day it mattered. Compliance tooling automates the configuring, and the proving is still a person with a folder of screenshots.

    GRC · Audit evidence · Control mapping · DPDP Act · Continuous monitoring

    Working version: Certinal

    20 min read

  4. How a lending decision gets warranted, not just made - a credit memo automation teardown

    Making the number is the easy half. Warranting it, with reason codes a regulator accepts and a trace nobody can quietly edit, is the half that decides whether the system ships.

    Commercial lending · ECOA · Model risk · Audit trail

    Working version: Arbiter

    19 min read

  5. What actually stops an agent is code it cannot reach - tearing down two agent prototypes

    Every governance feature that lives inside the model's context window is advisory. The test for any agent control is one question: can the model change this outcome by saying something?

    Agent governance · Policy gates · Prompt injection · Evidence ledgers · Human oversight

    Working version: Praxis

    19 min read

Every teardown here has a working version you can open. See the products