I scrutinize AI and data systems until the misleading numbers can't hide.
Most people report a model's best result and stop there. I stress-test mine until I find where they're lying to me — and I show you exactly where that happened, not just the number that survived.
A model looked like it had a real edge. It didn't. Here's the exact moment that became visible.
Political science degree. Self-taught in everything after that.
I didn't come up through a CS program. I came up through a warehouse-hot apartment, a wall of GPUs, and a market that changed its mind overnight.
I have a bachelor's degree in political science and history. Nothing about my formal education pointed toward machine learning or systems engineering — the technical side came entirely from building things myself, at scale, and living with the consequences when they broke.
That started with cryptocurrency mining. I built out a 30-GPU rig that, at its peak, was generating around $200 a day. It ran out of a small apartment in St. Louis — one room, no dedicated cooling, running at roughly 120°F most of the time. I had one monitor, so keeping the whole rig alive meant physically carrying that monitor from machine to machine, constantly restarting, constantly fixing.
It felt less like a data center and more like the final scene of a heist movie, if the vault were also a sauna.
Then Ethereum moved to proof-of-stake. Overnight, the economics that justified the entire rig stopped existing. Hardware that had been printing money became, functionally, a pile of expensive space heaters. That collapse is the actual origin of the way I approach every system now — I'd watched a number that looked completely reliable evaporate the moment the underlying rules changed, and the only reasonable response was to accept that verdict and rebuild, not to keep insisting the old numbers still meant something.
I didn't walk away from hardware — I redirected it. The next chapter was building and selling gaming PCs on eBay, which kept me hands-on with real builds and real customers. I kept the last few GPUs from the mining days, and that's what led into inference work: running genuine mixture-of-experts setups, splitting models across multiple GPUs working in sync rather than treating a single card as the whole system.
That's also where my specialty comes from — getting AMD and Nvidia cards to cooperate inside a single machine, split across two operating systems through WSL2. Most people building AI systems never have to solve this, because they rent a single-vendor cloud instance and the hardware problem is invisible to them. I solve it because I built my way up from consumer hardware that didn't want to cooperate, and had to make it work anyway.
From mining rig to raw-first AI systems
30-GPU mining rig, St. Louis
Peak output around $200/day. Constant hands-on maintenance, one monitor moved between machines, no real cooling — the hardest possible introduction to running infrastructure at scale.
Ethereum's move to proof-of-stake
The economics the entire rig depended on disappeared overnight. Hardware that had been profitable became unusable for its original purpose. First real lesson in accepting a verdict instead of arguing with it.
Gaming PC builds, sold on eBay
Repurposed the hardware and the hands-on build skill into a real, running small business.
Mixture-of-experts inference, mixed-vendor GPUs
Salvaged GPUs became the basis for real multi-model inference systems — AMD and Nvidia cards working together in one machine, split across Windows and WSL2.
AVAI — raw-first AI & data systems
Applying the same discipline to a real prediction system, Arena Racecourse USA — see the case study.
A prediction system built on one rule: nothing gets hidden, including the mistakes.
Arena Racecourse USA is a horse-racing prediction system built on a "raw-first vault" philosophy — every API response is saved byte-for-byte before anything is parsed or interpreted. That discipline is what made both of the findings below possible to catch.
Finding 1 — the lucky-winner trap
An early backtest of the GB model showed a genuinely striking result: +24.7% ROI against real Betfair starting prices across 10,489 bets. Good enough to get excited about. But breaking down where the profit actually came from told a different story — the five single largest winning bets contributed $4,648 of the total profit, including horses that paid out at 1000/1 and 191/1. That's not an edge, that's a small number of lottery tickets landing.
Gross (as reported): 10489 bets ROI +24.7% Top 10 winning BSPs: [1000.0, 1000.0, 191.3, 160.0, 100.0, 90.0, 85.0, 80.0, 80.0, 73.9] Profit from top 5 winners alone: $4,648
Restricting the strategy to a disciplined price band (BSP 3–20, cutting out the extreme long shots) brought the headline number down to +9.1% ROI — a smaller, far less flattering number. But the top 5 winners now accounted for just $180 of $1,136 total profit. The result got less impressive and more real at the same time. That trade is the whole point: a number that survives scrutiny is worth more than a number that doesn't.
Finding 2 — the bug that made the numbers look better
The second finding is the one worth remembering. A jumps-racing strategy, restricted to that same disciplined price band, was walk-forward tested across 10 months and came back looking close to perfect:
The dangerous bugs aren't the ones that crash. They're the ones that make your metrics look better.
A crash tells you immediately that something is wrong. A silent labeling leak does the opposite — it produces a cleaner walk-forward curve, more consistent monthly profits, a more "textbook" result. Nothing about the output invites suspicion. The only way to catch it is to already have a habit of interrogating results that look too good, and a raw vault detailed enough to trace exactly why they looked that way.
The model didn't survive walk-forward — so we're feeding it richer data and retesting
A negative result isn't the end of the story, it's a diagnosis. The original vault only ever stored settled, final prices. It had no record of how the market moved in the minutes before each race — and that movement is often where the real signal lives. That data can't be bought after the fact. It only exists if you record it while it's happening.
So on July 4, 2026, a dedicated tick recorder (feeder_ticks.py) went live, pulling GB bookmaker odds and US tote pool data every 60 seconds and writing every tick straight into the vault, byte-for-byte, the same raw-first discipline as everything else. Below is a real snippet of its own console output — not a mockup:
Snapshot of feeder_ticks.py's actual console output, recorded July 5, 2026. Shown here as a static excerpt, not a live feed.
The vault today
The plan
Let the tick recorder run for roughly a month, building a genuine line-movement dataset that didn't exist before. Then retrain and walk-forward test again — the same honest process shown above, run on richer data, with no assumption that it'll turn out any better.
A glimpse of Arena Racecourse USA's live race view
The case study above is the data honesty story. This is the actual product it feeds — a real-time race day dashboard, styled the way it runs in production.
Illustrative sample race, styled to match Arena Racecourse USA's actual live dashboard — badges, odds-movement arrows, and card layout mirror the production app. Swap in a real exported race card any time to make this live data instead of a mockup.
Draw your roof, pool, or lawn. Get a real quote.
A different domain, same instinct: turn a real-world shape into a number you can trust. Search your address, trace the outline, and this calculates real area and a real price live in your browser — nothing here is faked or pre-scripted. Then drop into Street View and measure real features against a single known reference.
This runs entirely in your browser — nothing you draw or measure is saved or sent anywhere. Estimates are intentionally rounded high and are illustrative only; official measurement and pricing always come from the contractor.
What the quote tool does — and what it deliberately doesn't
What it does
Search any address, trace the roof, pool, or lawn directly on the satellite image, and the tool computes the real measured area and prices it live in your browser. Drop into Street View and the AVAI Measure panel appears: click both edges of one object you know the size of — a window you've measured, a standard entry door — and that becomes a reference scale for estimating anything else on the same wall. No laser, no ladder; one known dimension in the frame is enough.
What it deliberately doesn't do
This is a demonstration, and it's built to be honest about that. Every estimate is intentionally padded high, so the official number should only ever come in at or below what you see here. Street View measurements are reference-scaled approximations — dependable on the same wall and at the same distance as the reference, not across the yard. Nothing on this page replaces a contractor's official on-site measurement, and no figure it produces is a quote or an offer.
Where this goes next
The interesting part isn't the widget — it's the feedback loop it makes possible. Every demo estimate paired with a contractor's real quote is a labeled data point, and over time that dataset teaches the system its own error: which job types it over-pads, how roof pitch shifts cost, where satellite tracing runs tight. It's the same methodology as the racing work above — log the prediction, wait for ground truth, measure the gap, correct — applied to something as ordinary as a roofing quote. That's what AVAI builds: systems that get more honest with use, not less.
The same pathology shows up anywhere a rate is computed from a filtered set of records
Sales & marketing
A conversion or close-rate report that quietly excludes "unqualified" or "bounced" records before computing the denominator can make a channel or rep look far more effective than they are — for the same structural reason a non-finisher leaking into training data made a model look profitable.
Ops & support
A resolution-time or SLA dashboard that drops "abandoned" or "reopened" tickets before computing the average will always look better than the real operation — and because the exclusion feels like data hygiene rather than distortion, it's rarely questioned.
In every case, the fix is the same discipline used above: keep the raw records intact, make the filtering logic explicit and inspectable, and treat any result that looks unusually clean as a reason to check harder — not a reason to stop looking.
What's in motion at AVAI
Short updates on what's being built, studied, and tested — posted as it happens, not as marketing.
Expanding into AWS & Google Cloud services
AVAI's systems have always run on self-managed infrastructure — a deliberate choice, and this site itself is served from AVAI's own hardware. But plenty of client work lives in the cloud, and AVAI is expanding its service offerings to include AWS and Google Cloud console work: deployments, data pipelines, and the same raw-first discipline applied to cloud-hosted systems. More on this at the end of the month.
Line-movement dataset: collecting, on schedule
The tick recorder described in the case study has been running since July 4, capturing GB bookmaker odds and US tote pool movement every 60 seconds. Five days in, we ran a full audit of our own collector — coverage verified against the racing calendar, one recording-density anomaly found, interrogated, and documented. The plan hasn't changed: roughly a month of collection, then retrain and walk-forward test on the richer data. Check back at the end of the month — the result gets published either way, good or bad. That's the whole point.
One standard, applied at every scale
Whether the job is a full prediction system or a slow laptop, the approach doesn't change: tell you what's actually true, deliver something complete, and never sell you more than you need.
Custom AI & data systems
End-to-end pipelines built on the same raw-first discipline as Arena Racecourse USA — data harvesting, vault construction, model training, and honest validation, including walk-forward testing that's allowed to fail.
In plain terms: I take large datasets — sometimes years of history — and train models on them to predict things like sales, demand, or customer behavior. Just as often, the process itself turns up where your data shows real waste: gaps, mislabeled records, or blind spots that are quietly costing you money.
- Local and multi-provider LLM orchestration (mixture-of-experts across mixed AMD/Nvidia hardware)
- Data pipelines with raw-first storage and explicit, auditable filtering logic
- Predictive modeling with real backtesting — including reporting when a result doesn't survive scrutiny
WordPress & web builds
Full sites built and configured properly the first time — clean structure, real performance, no unnecessary plugin bloat.
Because I work directly at the code level — not just inside a drag-and-drop builder — nothing about your site is locked into a template. Any element can be customized down to the raw HTML, CSS, or PHP, which means genuinely unlimited flexibility instead of "whatever the plugin allows."
- Business and portfolio sites
- Self-hosted deployments when that's the better fit (this site runs on my own hardware)
PC cleanup & optimization
Most home computers are slower than they need to be because of software that made sense ten years ago and doesn't anymore.
- Removing bloatware and unnecessary paid security suites (e.g. legacy Norton/McAfee subscriptions) when built-in, ISP- and browser-level protection already covers typical home use
- Startup cleanup, junk removal, and general speed recovery
- A plain explanation of what your machine actually needs going forward
The same rule at every price point
- If a system is broken, I say so — not a vague "let's monitor it," an actual answer.
- If something you're paying for isn't earning its keep anymore, I'll tell you, even when it means less for me to sell.
- Every deliverable is complete and usable on delivery, not a partial fix that needs a follow-up visit.
- If a result looks unusually good, I check it harder before I hand it to you — not after.
Tell me what you're working with
Whether it's a full AI system, a website, or a computer that's just gotten slow — start with what's actually going on, and go from there.
Terms of Use
Last updated: July 2026
These Terms of Use govern your access to and use of this website, operated by AVAI (Alex Vance Artificial Intelligence), a service of AVAI LLC ("AVAI," "we," "us"). By using this site, you agree to these terms.
Use of this site
This site is provided for informational purposes to describe AVAI's services and past work. You may view and share its content for personal, non-commercial reference. You may not reproduce, redistribute, or republish substantial portions of this site's content, including case study text and data, without prior written permission.
No professional, financial, or gambling advice
The Arena Racecourse USA case study on this site describes a personal research and modeling project. It is presented to illustrate a data-methodology approach and is not investment advice, financial advice, or gambling advice, and it is not an offer or solicitation to place wagers or trade any financial instrument. Historical performance figures shown are backtested results on historical data and are not a guarantee or prediction of future results. Sports wagering involves financial risk; if you choose to wager, do so only through licensed operators in your jurisdiction and within your means.
Property quote demo
The instant property quote tool on this site is a demonstration only. The estimates it produces are intentionally padded high and are approximations — not quotes, offers, or binding prices. Street View measurements are reference-scaled visual approximations for illustration. Any actual work requires an official on-site measurement and a written estimate from the performing contractor. The demo runs entirely in your browser; nothing you draw or measure is transmitted to or stored by AVAI.
Intellectual property
The AVAI name, logo, and site design are the property of AVAI LLC. Code, data summaries, and written case-study content are provided for reference only and remain the property of AVAI unless otherwise agreed in writing.
No warranty
This site and its content are provided "as is" without warranties of any kind, express or implied. AVAI does not guarantee the site will be error-free, uninterrupted, or that any described results are reproducible in every circumstance.
Limitation of liability
To the fullest extent permitted by law, AVAI and its owner shall not be liable for any indirect, incidental, or consequential damages arising from your use of this site or reliance on any information presented on it.
Changes to these terms
These terms may be updated from time to time. Continued use of the site after changes are posted constitutes acceptance of the revised terms.
Contact
Questions about these terms can be sent to Alexjdvance@avai.llc.
Privacy Policy
Last updated: July 2026
This Privacy Policy explains how AVAI (Alex Vance Artificial Intelligence / AVAI LLC) handles information in connection with this website.
Information we collect
This site is a static site and does not use cookies, tracking pixels, or third-party analytics by default. The only personal information AVAI receives is what you choose to send directly — for example, if you email us using the contact link, we receive your email address and whatever information you include in your message. The property quote demo runs entirely in your browser: addresses you search, shapes you draw, and measurements you take are not transmitted to or stored by AVAI.
How we use information
Any information you send us directly (such as via email) is used solely to respond to your inquiry and, if you engage AVAI's services, to deliver those services. We do not sell, rent, or share your personal information with third parties for marketing purposes.
Third-party services
The property quote demo loads map, Street View, and address-search functionality from Google Maps Platform. When you use it, your browser communicates directly with Google, whose collection and use of that data is governed by Google's own privacy policy, not this one.
Hosting
This site is self-hosted on infrastructure operated by AVAI. If that changes to include third-party hosting, analytics, or advertising services in the future, this policy will be updated to reflect what data those services collect.
Data retention
Emails sent to AVAI are retained as needed to respond to your inquiry and for standard business record-keeping, and are not used for any purpose beyond that without your consent.
Your rights
You may request that AVAI delete any personal information you've previously sent us by emailing the address below. We will honor reasonable requests except where retention is required for legal or legitimate business record-keeping purposes.
Changes to this policy
This policy may be updated from time to time. The "Last updated" date above will reflect the most recent revision.
Contact
Questions about this policy can be sent to Alexjdvance@avai.llc.