Chandrakanth Thadkapally

I have spent over a decade building the systems that move money and records between organisations — retail commerce platforms, payment reconciliation, healthcare staffing, HCM and payroll, telecom provisioning. Different industries, one recurring shape: two systems believe different things about the same event, and something has to decide which belief is right.

For most of that time the deciding was done by rules, and by people working the cases the rules could not. Rules are excellent at what somebody anticipated. They degrade quietly on what nobody did, and the degradation does not announce itself — it arrives as a slowly growing exception queue, and the size of that queue becomes the cost of the system rather than a signal about it.

What has changed is that we can now put a model in that seat, and increasingly an agent: something that plans, calls tools, and acts without a human between each step. That is the most interesting engineering problem I have worked on and the one with the least settled practice around it.

Most of the attention goes to the model. Almost none goes to the ring of code around it — the tool surface, what gets into the context window and why, the control loop, what happens when a call fails. I have come to think that ring, the harness, explains more of the variance between two teams’ results than the choice of model does. It is also the part nobody designs deliberately.

The second gap is verification. An agent that is right most of the time is a demo, not a system, and in most real work there is no gold label to check it against — two experienced reviewers disagree, so accuracy is undefined before you start. The third is control and confidentiality: what did the system disclose, and what evidence exists that it operated correctly. Data masking answers neither. An AI policy answers neither.

I work these questions out in payments and reconciliation, because it is the domain least willing to accept a confident guess. My current work at Walmart Global Tech is payment reconciliation and retail correction at scale — event-driven services, financial data pipelines, and the analytics surface finance uses to confirm that transactions are accurate. It produced two manuscripts, ReconGraph and NeuroRecon, both complete and neither submitted anywhere yet.

The breadth matters more than it looks. A commerce catalogue, a staffing platform, a payroll run and a settlement file are the same problem wearing different clothes, and having built all four is what makes the pattern visible. What generalises is the shape of the failure, not the schema.

Everything on this site uses synthetic data, describes classes of problem rather than any specific employer’s systems, and states what it does not know. If you work on any of this, I would like to hear from you.