Chandrakanth
Thadkapally
Enterprise AI agents for systems that have to be right.
I build enterprise AI agents, and the architecture that makes them safe to run. Twelve years across retail commerce, payments, healthcare staffing, HCM and telecom taught me the same lesson each time: the interesting engineering is not the model, it is the ring of code around it and the evidence it leaves behind.
This site is the working record — essays under The Control Loop, the systems I have built, the research programme behind them, and the factual record.
Plan, act, observe, verify. The loop only earns trust at the gate, and the branch that matters is the one that fails it — what an agent does when it is wrong is the whole design.
- Role
- Senior Technical Lead
- Based in
- Bentonville, Arkansas
- Focus
- Enterprise AI agents & architecture
- Open to
- Collaboration, review, speaking
The Control Loop
- 25 August01
Why Data Masking Is Not Enough for Enterprise AI
Masking removes the identifiers. It does not remove the structure, the logic, or the accumulation — and in an enterprise, those are the three things you actually cannot afford to hand over.
Enterprise AI privacy11 min - 25 August02
Reconciliation Is a Graph Problem, Not a Matching Problem
Pairwise matching assumes every obligation is settled by exactly one movement. Real settlement splits, aggregates and lags, so the assumption fails on precisely the population that costs money.
Payments & reconciliation15 min - 25 August03
Seven Ways Deterministic Record Matching Breaks at Enterprise Scale
A matcher that clears 97% of items is not 97% right. The residual is seven recurring structural classes, and every tolerance rule added to absorb one buys silence rather than accuracy.
Payments & reconciliation20 min - 25 August04
Neural Models Do Not Respect Accounting Identities
A model can fit every component of an allocation and still fail the only test that matters — whether the parts sum to the whole. Four places to enforce that, and what each one costs.
Constrained & neuro-symbolic ML16 min
The systems that most need machine learning are the ones least able to tolerate a plausible answer.
Agent engineering
The harness around the model — tool surfaces, context assembly, control loops, and how you evaluate an agent when no gold label exists.
Governed agentic AI
Making agent behaviour auditable where it matters: least-privilege context, approval gates that are not theatre, and evidence a control tester will accept.
Enterprise AI privacy
Confidentiality failures that survive masking: relationship leakage, rule leakage, and disclosure that accumulates across sessions rather than within one.
Payments & reconciliation
Transaction correspondence, settlement and lineage treated as problems over heterogeneous temporal graphs rather than pairwise string matching.
Constrained & neuro-symbolic ML
Models that must satisfy hard identities — conservation, balance, double entry — and the four places a constraint can actually be enforced.
Retail & commerce systems
Applied write-ups from production-scale commerce platforms: search architecture, catalogue retrieval, and the parts of a retail stack that carry load.
- RGReconciliation as a graph problemIn progress
- NRLearning under accounting identitiesIn progress
- CFLeast-privilege context for agentsIn progress
- ReconGraphIn preparation
Probabilistic Consistency Inference over Heterogeneous Temporal Graphs for Enterprise Financial Reconciliation
Enterprise reconciliation is not a record-pair problem. A single economic event can produce an order, an invoice, a shipment, a receipt, a payment, a settl…
- NeuroReconIn preparation
Constraint-Aware Neuro-Symbolic Learning for Enterprise Financial Reconciliation
Enterprise reconciliation requires more than recognising that two records look alike: an accepted match must also satisfy monetary conservation, event orde…
If your systems have to be right, I would like to hear about it.
I read everything sent to me about agent architecture, evaluation without ground truth, and putting autonomous systems into places that get audited. Review requests, collaboration and speaking enquiries are all welcome — and disagreement from people who have run this in production is the most useful mail I get.
Get in touch