In every regulated institution, the SOP, the floor, and the audit trail quietly diverge. We make them one artifact — prose for the auditor, code for the machine, one spine underneath.
# the same sentence, executable artifact "your-process": skins: [prose, code] drift: mechanically_detectable controls: cannot_be_waived trust: earned_per_run
Every operations leader in financial services knows the secret: the process manual describes a fiction. The floor runs on workarounds, tribal knowledge and judgment the SOP never captured. The few places where documentation matches reality are usually the places where the documentation was written after the fact, to mirror what people were already doing.
This gap — not bad intent — is the root cause of repeat audit findings, of training that doesn't train, of controls that exist on paper while the floor routes around them.
AI agents make it possible to close the gap, but only inside one discipline:
One source, many projections, drift mechanically detectable.
We call it the single-spine principle. Everything we build hangs off it.
A ProcessSkill is a single document in which the SOP your auditor reads and the code your execution engine runs are the same artifact. Edit the prose and the code must follow; delete a control point from either side and the document refuses to deploy.
Markdown the regulator can print; executable nodes the harness can run. Drift between the two is not reviewed annually — it is detected mechanically, on every change, before deployment.
Every process ships with a validator specification — authored independently by the control function, builder-blind. Fluent, self-consistent output cannot reward-hack an oracle it has never seen.
Judgment-heavy exceptions — the transliteration variance, the unregistered partition, the disputed dues — become executable under the same governed spine, instead of falling to a queue.
At regulatory gates the human decides on the case file first; the agent's recommendation stays sealed until the decision is logged. Divergence becomes a third-line signal, not an argument.
Yantra is our process-orchestration harness: a thin deterministic boundary around capable but unreliable models. model proposes, harness disposes — agents act only inside spine nodes, every checker's independence is enforced at dispatch, and the run's event log is the audit trail. The regulator pack, the customer's status, the cost-per-decision dashboard: all projections of one record.
We keep the harness deliberately thin, on commodity orchestration substrate, designed against RBI's FREE-AI framework — because the defensible part of this work was never the engine. It is the artifact, the wall, and the trust they earn per run.
We mine your event logs — LOS, LMS, CBS — and put the as-executed process beside the official SOP. Every divergence, quantified and ranked. You leave with the founding artifact of any honest transformation: a map of how work actually happens.
One process, authored as a ProcessSkill with an independent validator, run on the Yantra harness for ninety days — in shadow or maximally gated. Three numbers decide what happens next: conformance rate, cost per decision, audit-pack production time.
The full blueprint: process-by-process migration on the strangler-fig pattern, knowledge canon and memory architecture, role transitions from maker to checker, governance mapped onto your existing committees — never a parallel structure.
Each template is a working starting point for a real NBFC process — policy pins, control points, checker classes and human gates included. The Process skin is open for anyone to read. The full file, Skill skin included, is sent on request.
Name and address mismatches, transliteration variance, re-work loops — the long tail of onboarding, structured as agent steps with a deterministic gate on every pass and a human decision on every exception.
Read the Process skin →Chain-of-title analysis with maker and checker agents of enforced different lineage, evidence floors, and decide-then-reveal at sanction. The highest long-tail value in secured lending.
Restructure options drafted by agents strictly within board-approved hardship principles — every offer citing pinned policy clauses, every commitment behind a human gate.
Alert triage with assembled evidence trails and draft STR narratives — the analyst owns the filing decision, the run record owns the defensibility.
Weekly writing on process governance, AI agents in regulated finance, and the gap between what institutions document and what they do.
Why the gap between the written and the executed process is structural, why twenty years of BPM couldn't close it, and what changed.
On validator independence, correlated model failures, and why maker–checker needs rethinking when both are machines.
No product announcements, no noise. Unsubscribe anytime. Template unveilings ride along with the column.
ProcessSkill is a transformation practice focused on RBI-regulated NBFCs. We sit at the intersection of process design, AI architecture and Indian financial regulation — DPDP, PMLA, and the RBI's FREE-AI framework are design inputs here, not afterthoughts.
We are deliberately small, deliberately opinionated, and we publish our methods. If the writing on this site convinces you, the conversation will be short.