About
Isaac Mineo — Senior backend engineer
- Role
- Senior backend engineer
- Location
- New Haven, Connecticut
- Middlebury
- 2019 to 2023
The short version
I build AI products end to end, and I have shipped five of them alone: two on the App Store, one trading real money unattended since April. At Accenture's Innovation Garage I have contributed to over seventy proofs of concept, each one starting from a business problem somebody could describe but not specify. I am most useful where the requirements do not exist yet and somebody has to sit with the customer and work them out.
70+
proofs of concept with enterprise clients
5
products of my own, two on the App Store
19
repositories, across 169 days with a commit
138
days arbor has traded unattended
Open to forward deployed and applied AI engineering roles.
Neuroscience, then systems
I read neuroscience at Middlebury, 2019 to 2023, and the computational half is not a stretch: Neural Coding, Data Science in Neuroscience, Data Structures, Computer Architecture, Linear Algebra, Psychological Statistics. I graduated in May 2023 with a 3.62, Dean's List and College Scholar.
The part that stuck was not the biology. It was that a model of a person is a claim, and a claim has to be checked against what the person actually did, or it is a story you are telling yourself with extra steps.
That is the same problem I work on now, wearing different clothes. KinoLog commits to a number before you watch a film and then measures itself against the rating you type. Bodivo will not state a trend it cannot evidence. arbor grades its own conclusions against the exchange ledger rather than against its own logs. None of that is a machine learning technique. It is the habit of building the check in at the same time as the thing being checked, which is what four years of being told that an unfalsifiable model is not a model does to you.
What the Garage ships
I joined Accenture on 22 July 2024 as an Analyst and was promoted to Senior Analyst in December 2025. My official title is Technology Senior Analyst, Application Engineering; the work itself is senior backend engineering. I am based in New Haven and onsite in Hartford.
The Innovation Garage does client-embedded rapid prototyping. An engagement starts with a client who has a problem they can describe but not specify, and ends with something running that they can put in front of their own people. The work in between is mostly not typing: it is sitting in the room, finding out what the constraint actually is, and being wrong in public early enough that it is cheap.
Building inside somebody else's context is the part that does not transfer from personal projects. Their data model, their compliance posture, their procurement, their two incompatible internal systems that both claim to be the source of truth. You do not get to choose the stack, and the thing you ship has to survive being handed to a team that did not build it. The pace is weeks, not quarters, which decides most of the architecture on its own.
One concrete thing, because a count says how often and not what. I built and shipped an internal deployment platform that lets teams deploy applications behind a security perimeter: an application load balancer distributing traffic, ECS Fargate orchestrating the containers, reverse proxies in front, Cognito user pools and groups for authentication, and AWS Verified Permissions for fine-grained role-based authorization. I carried it from staging through to production. Describing what I built is not describing who I built it for.
Seventy repetitions teach something that shipping your own products does not. On your own work you can change the problem when it gets hard. On somebody else's you cannot, so you learn to find the actual constraint fast, because the alternative is discovering it in week three. You also learn that the demo is the deliverable: a prototype nobody can follow is a prototype that does not get funded, and being able to explain it to the person paying for it is not a soft skill, it is the job.
May 2023 to July 2024
I signed with Accenture in December 2022, six months before I graduated. Then the start date moved: Q1 2023, then September, then July 2024. Consulting firms deferred an entire campus intake that year and it was widely reported at the time. It was not about me and there was nothing to do about it.
So I spent fourteen months at home. My nephew had just been born and I was there for it, which is not a line I would trade for a faster start date. I worked retail while I waited. If the dates on my resume raise a question, that is the answer, and I would rather write it here than have somebody guess.
Shipping alone, on purpose
Client work ends at the handover. I wanted the other half: the schema I chose in month one still being the schema in month nine, the support email from somebody who is annoyed, the review rejection that is technically correct. Five products, built end to end and alone. Two are on the App Store, which means somebody at Apple has read them.
I am AWS Certified AI Practitioner and working through the Solutions Architect Associate. The certification came after the work rather than before it: the deployment platform in section 02 is Fargate, ALB, Cognito and Verified Permissions, in production.
That loop changes how you write a backend. App Store review taught me that a rule you disagree with is still a rule, and arguing costs a week. The support inbox taught me that the failure people report is almost never the failure that happened. And running a system that trades real money unattended taught me the thing I would put above all of it: a guard that logs and continues is not a guard. It has to halt, or it is advice.
How I decide
Five opinions, each earned somewhere
Deterministic where the answer is arithmetic. Generative only where it is language.
KinoLog predicts your rating with a nearest-neighbour model over your own diary, then hands the finished number to a language model and lets it write the sentence. It is never allowed to produce the value. RecipeFix rescales quantities in code and asks the model only to rewrite the method. This site answers questions over 310 chunks with BM25 and cosine in memory, because a vector database would add a network hop, a key, a cold start and a bill to replace a loop over an array.
Degrade honestly rather than plausibly.
Bodivo returns the reason a view is empty and what would fill it, rather than rendering an unknown as zero. This site is built to print [TK] in public for anything it does not know, because a visible hole beats a plausible fabrication in front of people who check. When no model key is set, the ask endpoint returns the retrieved passages verbatim with real citations: a worse-written answer and an equally true one. A function that hashed repository names into plausible commit-activity bars was deleted for the same reason.
Allowlist over blocklist, because a blocklist fails open.
Every publishing flag on a private repository defaults to false, so each published field is a decision somebody made rather than a default nobody reviewed. The retrieval corpus reads an explicit list of documents rather than a directory. The failure mode of a blocklist is that something new leaks; the failure mode of an allowlist is that something is missing, and missing is visible.
A system that never learns whether it was right is a toy.
I wrote that into a build specification two years ago and it has decided more architecture than anything else I have thought. KinoLog commits to a rating before you watch, settles it against what you actually typed, and feeds the signed error back as a bias correction that is shrunk toward zero, capped at four tenths of a star, and silent below six settled films: deliberately underpowered, so it can never do much damage. Bodivo classifies every metric as measured, calculated or unknown and refuses to state a trend below its sample threshold. This site's verify:ask fails the build when a suggested question retrieves nothing. None of the three is an evaluation framework in the benchmark sense. All three commit to something falsifiable before they find out, and bound how much the answer is allowed to change them.
A check that has never failed has never been tested.
Both of this site's verify commands were proved by planting a violation and watching the build exit 1. The first plant passed, which was itself the finding: the assertion had been defined and never called. A contrast audit across every route reported zero failures until it turned out to be reading oklch coordinates as if they were RGB and measuring everything at 1.00:1. Rasterised properly it found 32.
A model of a person is a claim, and a claim has to be checked against what the person actually did, or it is a story you are telling yourself with extra steps.