Agentic AI, operationalized with rigor.
Kalmi gets AI workflows into production at Indian and global financial institutions, and keeps them there. One workflow at a time: the work redesigned, controls built in so risk and compliance can sign off, adoption driven, and value measured.
The pilot works. The system still does not ship.
Some pilots fail on accuracy, on data that was never as clean as the demo, or on a weak business case. This is about the other ones: systems that work in test, have a sponsor and a budget, and still do not reach production or move the numbers.
Part of the gap is the work around the system. The process it was dropped into did not change: the same roles, the same decision rights, the same exception handling. Most pilots stall there rather than in the model.
The rest is control. The policy exists as a document no runtime reads. The logs were built for debugging, not for a reviewer reconstructing a decision a year later. The controls are described and unowned in production.
Risk and compliance are not being obstructive when they decline to sign. They are being asked to accept a system they cannot inspect, constrain or evidence.
One workflow at a time, in five fixed-fee phases.
Each phase is scoped and priced before it begins, and the engagement can stop after any of them without leaving you with half a system.
Triage and value case
Two to three weeks. We choose the one or two workflows that change the economics of the work, baseline the metrics they will be judged on, and name the use cases that should not proceed, with the reason. Most of an AI roadmap should not be built right now.
Workflow redesign
The target process comes before any tool: who decides what, where a person reviews, how exceptions queue and who owns them. A tool deployed into an unchanged process rarely moves the numbers. The minimum controls are specified here, not added after go-live.
Deploy
Tools chosen from a disclosed shortlist, or built where nothing fits, then integrated and taken through a go-live gate your risk function signs. Where a vendor's own team deploys its product, we work from your side: acceptance criteria, the controls, and everything around the tool.
Adoption and change
Training, operating procedures, an override policy, and incentives that match the new process. Most deployments stall here, so adoption has its own scope and fee rather than a handover session at the end.
Measure and operate
Value reported against the triage baseline: turnaround, rework, override rates, cost per case. We operate the workflow under a defined support window, with exit terms agreed at the start, or hand it to your team. Then the next workflow.
Where defensibility has to come first, the same work can start from a diagnostic: an inventory of the AI already in use, including vendor AI, the gaps against the frameworks that apply to you, and a remediation plan that names who operates each control.
Every workflow ships with the controls that let risk sign it.
Governance is not a separate project you have to buy first. It is designed into each workflow and enforced while it runs.
Minimum controls, in every workflow
An entry in the AI register, decision logs built to an evidentiary standard, human review thresholds, a kill path that stops the system, and value metrics in business terms. Present from the first deployment and priced inside it.
Policy enforced in the execution path
Where AI traffic runs through a gateway, each obligation becomes a rule evaluated before the action completes. The control fires or it does not, and either outcome is recorded with the policy version that applied.
Evidence over AI you did not build
Vendor AI usually runs outside your gateway. Its outputs, logs and attestations are collected from your systems and the vendor's, checked against the same policies, and kept as one decision record your risk committee and examiners can read.
The full control plane comes in when the estate needs it: several systems in production, a high-risk decision such as credit, or a question from model risk or the board. It is assembled from existing gateway products you license directly, and deployed in your environment.
What production-ready means here.
A system is production-ready when it delivers commercial return while meeting all six of these, defensible under the frameworks that apply to your institution. Four out of six is a pilot with a launch date attached.
Explainability. A decision can be reconstructed and explained to someone who was not there, in the language the business uses.
Transparency. What the system does, on what data, under which policy version, is visible without reading the code.
Consistency. The same inputs receive the same treatment, and drift is detected rather than discovered.
Correctness. Accuracy is measured against a defined standard continuously, not demonstrated once at launch.
Bias management. Assessed on the attributes that matter for the use case, with a documented method and a retest cadence.
Commercial return. The case still holds after the cost of running the system under the five constraints above.
This is the standard we hold client systems to, and the one our own work gets measured against.
Where we sit, and what we do not sell.
First line of defense
Kalmi builds controls and operates them, which places us in the first line. We do not provide second-line validation or third-line assurance on our own work, because the party that builds a control cannot challenge it. Your model risk function validates what we build. We build so that it can be.
Opinionated, with no commercial interest
We have a view on which products fit and we state it. We take no resale, referral fees, margin or vendor money, you contract with every vendor directly, and the shortlist we work from is disclosed, with at least two options per category.
Alongside vendor teams, not against them
When a vendor's team deploys its own tool, we do not compete for the configuration work. We work for you around it: the redesigned workflow, the go-live gate, adoption, and the value review ninety days later.
We decline work
Every triage names the use cases that should not proceed. We apply the same filter to ourselves: one engagement shape, a bounded set of named workflows, and no single-tool, low-risk deployment a vendor's own team will do as well.
Sudip Chakraborty
Sudip was a Vice President at Goldman Sachs, working on stress testing regulation, board-level governance and buy-side market risk: the discipline of making a model defensible to a supervisor rather than merely accurate. At Soroco he implemented ISO 27001 and SOC 2 compliance, which is control work rather than framework theory. IIT Guwahati, IIM Kozhikode.
This work requires the technical, the domain and the regulatory at the same time. Teams strong in two of the three produce frameworks that collapse in production, or systems that work and cannot be defended.
Two kinds of institution, two different reasons this stalls.
An NBFC supervised by the RBI
The pressure is to do more per loan with the same teams, while the model risk questions get sharper. AI that moves turnaround in credit or collections has to arrive with the inventory, the controls and the evidence you can produce on the day you are asked.
The India entity of a global financial institution
The build happens here and the model risk policy it has to satisfy was written elsewhere. A pilot clears internal review, reaches the edge of production, and stops because no one owns the redesigned process or operates the controls once it is live.
The first conversation is a diagnosis, not a pitch.
Bring a workflow that should be in production and is not, or a deployment whose numbers did not move. We will tell you what we typically see at that stage, specifically enough that you can confirm it or correct it. If an engagement follows, it starts with a paid, fixed-fee triage. We do not run free assessments.
Replies come from Sudip, usually within a working day.