Do I have a guarantee, or am I guessing?
Convergence analysis taught me to ask one question of any algorithm. That question shaped an agentic supply-chain system built to decide when not to trust its own answers.
Convergence analysis taught me to ask one question of any algorithm: do I have a guarantee, or am I guessing? That question shaped an agentic supply-chain system I built to decide when not to trust its own answers.
Why it matters
An autonomous agent that acts on a confident wrong answer can commit a company to an expensive decision before anyone reviews it. The safeguard is knowing which answers carry a proof and which are approximations, then gating the rest.
My research produced conditions an engineer can compute before training a network, rather than guessing through trial and error. That discipline, separating what is provable from what is merely plausible, is what production AI systems need most.
Chip Huyen’s AI Engineering makes the case that shipping a model and building a system are different skills. Andrew Ng frames an AI career around a portfolio of projects that shows skill progression. This system is where I put both to the test.
The credential that closes the gap is a system that knows the limits of its own answers.
How it works
The evaluation-versus-guess judgment runs through every layer.
Two solvers compute the optimum: linear programming for network flow, second-order cone programming for robust allocation under uncertainty. The rest return the best solution reachable on problems that are provably hard, and the system labels which is which rather than passing off a heuristic as optimal.
The safety layers apply the same test. A CRAG pipeline drops irrelevant documents before the model sees them. Typed contracts guard every tool boundary. 20 adversarial probes target at least 98% injection resistance in CI. Any routing decision above $10,000 stops for human approval before execution.
Intent classification follows the Stanford NLP “program, don’t prompt” principle through DSPy: declarative signatures compiled against the model replace brittle hand-written prompts, so swapping the underlying model recompiles the program instead of forcing a prompt rewrite.
What I didn’t expect was how little distance there was between the two worlds. The instinct that tells you a training algorithm will converge is the same one that tells you an agent’s answer needs an evaluation before it ships. The research was never separate from the engineering.