Leadership
Influence I can point at, and one claim I am keeping small
Three sections, kept separate on purpose. I lead technically and across teams, and I can show you where. I have not managed people, and nothing on this page is arranged to suggest otherwise.
Technical leadership
Setting technical direction, and being willing for that direction to be smaller or less flattering than the one requested.
Recommended less work than I was asked for
The brief was to build a model router. I measured the fleet first, found model choice accounted for roughly one percent of the bill, and recommended a routing middleware of about 150 lines plus cache hygiene instead of the routing programme that had been scoped. The hardest part of the work was arguing for a smaller version of it.
What a Production Agent Fleet Actually Costs →Four disagreements, recorded in the deliverable itself
The study documents where I disagreed with the brief rather than resolving it quietly: that the problem was budgets rather than models, that per-call routing had negative expected value against a measured switching cost, that the legacy evaluation suite must not gate anything, and that a one-day policy change outranked the engineering programme. Writing disagreements into the artefact is what makes them reviewable instead of remembered differently by each party.
What a Production Agent Fleet Actually Costs →Published a do-not-do list
Alongside the recommendations, an explicit list of things not to build, each with the measurement that ruled it out. A recommendation without its rejected alternatives is an opinion; with them it is a decision someone else can re-open on new evidence.
What a Production Agent Fleet Actually Costs →Audited my own safety net and published the negative result
Before letting a 400-scenario evaluation suite gate a cost programme, I checked whether its assertions bound to anything. 265 of 271 in the legacy layer asserted nothing. Reporting that made my own prior work look worse and made every subsequent decision sounder.
Enterprise NL-to-SQL Agent Platform →Set the standard for what a result may claim
A written retraction policy with a named owner, validity gates that derive a do-not-quote list from gate state, and a results changelog enforced by a test. Governance implemented as mechanism rather than as a norm people are asked to remember.
AI Benchmarking & Measurement Platform →Established failure-direction conventions the codebase follows
Fail closed on ambiguity, fail open on infrastructure. Applied consistently enough across guards, authorization, memory and routing that new contributors inherit it without being told.
AI Compliance Investigation Platform →Replicated a pattern until it became an architecture
Build the offline pipeline that makes the question cheap, then put a conversational agent on it as a domain in the shared platform. Four systems followed that shape, each built standalone and each designed to be absorbed. That repetition is the difference between building agents and defining how an organisation turns an analytics asset into a product.
Inventory Intelligence Platform →Review discipline, including on my own work
Two commits in the platform's history exist purely to remove unrelated changes from my own pull requests. A review standard that only applies to other people's changes is not a standard.
Enterprise NL-to-SQL Agent Platform →
Cross-functional influence
Changing what other teams do, using evidence they can check rather than authority I do not have.
An executive evidence pack with a no-unsourced-claims rule
Every figure in the pack carries its basis: verified in code, observed in traces, calculated from published prices, or modelled as a range. The rule was not decoration: it meant the numbers survived scrutiny in rooms I was not in, which is the only property that matters for a document that travels.
What a Production Agent Fleet Actually Costs →Named a one-day policy change as the top engineering priority
An organisational allow-list was blocking access to a model family with materially better economics. The highest-value item on the list was not code at all, but a ticket someone else had to file. Saying so cost me the more interesting recommendation and was the correct call.
What a Production Agent Fleet Actually Costs →Turned an unmeasurable baseline into a leadership ask
Some of what the study needed could not be measured with the instrumentation that existed. Rather than modelling around the gap, I specified what would have to be instrumented and asked for it explicitly, with the decisions it would unblock attached.
What a Production Agent Fleet Actually Costs →Wrote a library other engineers adopted
A multi-provider batch runtime I authored alone was picked up and vendored into other teams' projects, including one I barely contributed to. Adoption without advocacy is the version of influence that can be checked without taking my word for it.
Multi-Provider LLM Batch Runtime →Put domain logic where domain experts can review it
Structured knowledge files, skill packs and closed factor vocabularies instead of prompts. It changes who is allowed to correct the system: a compliance specialist or an operations analyst can read and change the rules they own, without an engineering ticket.
Carrier Root-Cause Validation Agent →Spoke to planners about forecasts, not to statisticians
A forecasting feature was technically correct and unusable, because it was explained in the vocabulary of the method rather than the vocabulary of the job. Rebuilding it around what a planner actually decides was worth more than any accuracy improvement would have been.
Enterprise NL-to-SQL Agent Platform →
Mentoring
Mentoring junior data scientists on eval-driven production AI.
Stated plainly, and not embroidered
I mentor junior data scientists on eval-driven production AI. That is the claim as it appears on my résumé, and I am not going to dress it up with specifics I have not written down. If it is relevant to a conversation, I would rather talk through what I actually teach: that the evaluation is part of the system, that a green suite you have not audited is a liability, and that the honest version of a result is usually the more useful one.