Tyler Brown

Applied AI & Systems Engineer  ·  Sales Manager, Innovative Bioscience

I carry a sales number. I build the systems behind it, and some of them are machines.

Over the last two years at a small life science distributor I have designed and shipped the systems a company that size is not supposed to have: a standards-compliant e-procurement connector that lets university buyers purchase from inside their own purchasing platform, retrieval-augmented search over the company's institutional knowledge, locally hosted language models doing extraction work at scale on one consumer GPU, and the guard layers that stop automated outreach from doing something that cannot be taken back.

The work is measured rather than asserted. 614,000 lines of JavaScript, Python and shell across twelve production systems, 824 commits across seven repositories, 643 of them in the main one, and 553 test files carrying 5,836 test cases and 13,267 assertions behind them. In the commit log, fixes outnumber features. Where a result did not hold up, I have said so in writing and withdrawn it.

Stack

GET punchout.innovativebiosci.com/punchout/health 200 OK 2026-09-09 06:21 UTC 
{
  "status":   "ok",
  "service":  "IBS PunchOut Connector",
  "version":  "1.0.0",
  "mongo":    "connected",
  "shopify":  "connected",
  "queues": {
    "needs_review": 0,  "failed":     0,
    "received":     0,  "processing": 0,
    "fulfilled":    0,  "outbound_dead": 0
  }
}
The connector described below, answering for itself. Both dependencies up, every queue at zero, no orders stuck or dead-lettered. The link is live if you want a fresher one.

Also running in public

Commit type · 599 of 643 main-repo commits carry a type prefix

fix224
feat206
docs100
test33
chore27
security9
Fixes outnumber features, 224 to 206. That is a maintainer's ratio, not a builder's. The remaining 44 commits use rarer prefixes or none.

Selected systems built
  • measured
  • verified
  • withdrawn

Applied AI and machine learning

Literature extraction on local GPU inference

A pipeline that reads the Methods section of published open-access papers, extracts which commercial supplier a laboratory actually used, and joins it to that author's published contact address. A 14-billion-parameter model runs locally through Ollama on one consumer GPU at zero marginal cost.

The model must return a verbatim source span, checked in code as a literal substring. A supplier name not present in the paper is discarded rather than reported. That is a structural hallucination filter, not a prompt instruction.

A truncation defect turned out to be the single cause of every malformed-output retry. Repairing the output instead of retrying it recovered 80 previously lost papers and produced zero failures across the following eight thousand.

THROUGHPUT · PAPERS PER MINUTE 7.4× 1.27 9.43 0 4 8 10
Quantization choice, not a bigger machine. Selecting a build that fits entirely in video memory did more than every other optimisation combined.

An evaluation-contamination catch on my own work

A prompt revision measured as a large quality gain. I then found the evaluation set had been assembled out of the previous prompt's own failures, with worked examples lifted verbatim from papers inside it.

Re-measured on randomly sampled, blind-labelled papers, the two versions were a dead heat, 21 of 24 correct each. I reported the tie, kept the revision on other grounds, and wrote the rule down: an evaluation set built from a model's failures will always flatter the fix aimed at them.

Retrieval platform, and the measurement of it

Company-wide semantic search over contracts, price sheets, product data and operational history, used by people and agents alike. Pinecone vector store, Gemini embeddings, MongoDB semantic cache and query audit log, contextual chunking in which every chunk carries a written summary of its own context.

I then built a falsifiable benchmark across every knowledge layer in the estate, and found my own earlier headline figure was inflated by answer leakage in the test harness. I voided it. The corrected result stands: a cross-layer fan-out answers 0.81 of questions from printed text against 0.42 for the vector store alone.

Generative media with a control layer

The hard part of generative media is not producing one good frame. It is the hundredth one matching the first, and the label still saying what the label actually says.

A product photography pipeline takes a phone photo in and returns a studio shot, behind an automated pass and fail gate with auto re-roll. The discriminator was measured rather than guessed: border colour and border texture carried no signal at all, reading identically on good and bad batches. What separated them was floor-band darkness, a 5th-percentile luminance of 52 on the bad batch against 90 to 159 on the good ones. The threshold sits at 90.

Label integrity is a separate guard that warps the original label pixels onto the generated bottle and fails closed: two of three generations died on "no label found" rather than shipping something plausible. Without it, at temperature zero, the fine print degraded into nonsense and the phone number changed by one digit.

The limits are documented as carefully as the wins. The gate cannot catch a wrong material or invented geometry, so a human pass on a contact sheet is still required. Real storefront photography measures at mean luminance 205 while the generated scenes came back between 55 and 137, up to four times darker and reading as a different company in feed. Realism and brand pulled in opposite directions, and brand won.

For video, type is burned in by code because the models garble text, a current video model cannot hold hands past about two seconds so hands get a still and a slow move instead, and scripted browser capture drives the cursor along Bezier paths with an eased clock, because evenly spaced replay reads as a robot.

Integrations and protocols

cXML and OCI e-procurement connector

The largest single system I have built. It lets an institutional buyer punch out of their own purchasing platform into our storefront, shop at contracted prices, and return a cart as a purchase order. Seven cXML message types plus SAP OCI, validated against vendored official document type definitions, with per-buyer pricing in four schemes applied on both cart return and order ingestion.

Built for the failure cases: a durable outbound delivery queue with at-least-once semantics, full-jitter backoff, strict document ordering and a dead-letter path; a stuck-order recovery sweep; atomic idempotency on purchase order intake. Two health endpoints are deliberately split: one always reports success for liveness, the other reports an honest failure when a dependency is down.

Nine rounds of adversarial security review closed more than forty findings, including an unauthenticated NoSQL operator injection leading to session hijack on the one service that creates real customer orders.

Marketplace order connector

A second protocol leg on the same service, receiving signed CloudEvents webhooks over HMAC-SHA256. It verifies the signature fail-closed, persists the raw signed event before doing anything with it, acknowledges inside the partner's three-second budget, then maps asynchronously. Money is handled as integer cents through BigInt, so float drift is not possible. Idempotency runs at two levels, event-level at the inbox and effect-level on the order, so at-least-once redelivery can never create a second order.

Two silent order-loss defects were caught before go-live. One of them, a signature header the partner names differently from the documented default, would have rejected one hundred percent of real orders while returning a plausible-looking response.

Live carrier rating at checkout

Replaced flat shipping tables with a callback service that rates every cart against a live negotiated carrier contract. It bin-packs against measured carton geometry, real stack heights and a 42-pound strapping limit, and forces cold-chain items onto next-day service, never bundled, with dry ice declared per package.

The governing design rule is refuse, never guess. If the carrier returns nothing usable, or any item lacks a weight, the service returns no rate rather than inventing one. A wrong rate is charged to a real customer. A missing rate is merely visible.

Safety architecture and data

A seventeen-layer pre-send guard

Not a mail sender. A gate that answers one question, may we contact this address right now, called by every drafting pipeline. Layers run in a fixed order and return on first block: format and domain checks, deliverability, provenance grading, suppression with alias resolution, per-person cooldown, draft deduplication, domain reputation drawn from our own bounce history, mailbox verification, three independent checks against a colleague's mailbox so the automation cannot double-contact his book, relationship tags, identity drift, and frequency and per-domain caps.

Fail-open and fail-closed posture is assigned per layer and documented as a deliberate asymmetry rather than left to chance. Two halts are independent by design: an operator stop that only a human may lift, and an automatic bounce watchdog that trips on a rolling rate.

The controlling finding came from audit, not from code. Every one of the 25 hard bounces the company had ever had was an unverified address. Verification, not volume, controls the bounce rate.

Buyer identification from public federal award data

A three-phase pipeline that scores every funded research project for whether that laboratory consumes what we sell, resolves the purchasing contact, and verifies deliverability before anyone is contacted.

The durable finding is that published-email yield is a property of a state's institutions, not of the crawler: California 27 percent, Texas 13 percent, Massachusetts 8 percent. Re-crawling every empty Massachusetts laboratory with improved extraction produced zero additional addresses, and a probe found zero bot-walls. That is a data limit rather than a tooling limit, and it redirected a whole territory from email to telephone.

Splitting the work so that paid inference does only the web search while local code does the fetch and deterministic extraction cost 84,000 tokens per batch against 387,000, about 4.6 times cheaper, and produced better coverage besides: 62 percent of subjects with a verified site against 11 percent.

Search, content and extraction

What the blog was actually doing

A 166-article technical corpus that carries 72.6 percent of all site impressions, 520,733 over three months, and converted at 0.47 percent. The measurement that mattered was narrower: 19,497 impressions at an average position of 6.7 with a 0.37 percent click-through rate. The corpus is read. It is not clicked.

I tested the two obvious fixes and both failed on measurement. A page carrying a working interactive calculator ran live for nine days with no change in click-through, because the search result cannot see a tool on the page. A title and description rewrite was ruled out after finding the same queries earned zero clicks even at position one with the price already in the snippet. The result absorbs the question.

So the work went into the conversion path instead. A changeset engine edits live articles and product pages against live anchors, dry run by default, writing a rollback file before any write and verifying every anchor afterwards. 31 of 31, then 11 of 11 operations verified live, every anchor matching exactly once and all 35 links returning 2xx.

Along the way: a theme defect where the article call to action was substring-matched on tags, so ppe matched charcoal-stripped and more than a dozen articles offered the wrong product family; and a full citation audit of the corpus that resolved 178 identifiers and found six dead source links, one bad identifier and 47 misattributions.

Extraction at scale, and what it taught

Reading the public web reliably turns out to be a de-obfuscation and provenance problem far more than a crawling one. The platform runs browser automation with fingerprint rotation across more than twenty user agents and eight viewports, per-domain adaptive backoff behind a circuit breaker, and eight extraction strategies ranked by confidence: mailto, JSON-LD, microdata, meta tags, labelled text, vCard, de-obfuscation, and raw pattern matching last. De-obfuscation covers HTML entities, reversed strings, ROT13 attributes and JavaScript concatenation. Decode, never guess.

The findings are worth more than the crawler. At one institution 92 percent of the database's guessed addresses were wrong, because the school issues opaque IDs rather than firstname.lastname, and ten of eleven sampled domains were catch-all, so verification was not the gate there at all. Published-address yield turned out to be a property of the institution rather than of the tooling.

Two defects worth naming. A bare @ counted as evidence a page had an address, so a CSS @media rule made JavaScript-heavy sites skip the headless render entirely and their rosters were never seen. And a pinned browser user agent quietly returned 144-byte empty documents.

One directory vein produced 1,014 published addresses from 1,198 pages with zero duplicates and zero missing provenance. A separate pass pulled 557 core-facility URLs across 30 institutions out of a public web index without crawling anything.

Machines and physical systems

A six-axis robot that prints buildings

Toolpath planning and machine control for a KUKA KR 210 on a six-metre linear track, printing concrete structures. It takes a parametric building model, plans the print path inside the arm's 3.1-metre reach envelope, solves for where the robot has to sit on its track to cover the whole footprint, and emits executable machine code with pump control.

The output is not a picture of a robot. One run produces 13,215 lines of G-code across 200 layers: 8.8 kilometres of toolpath, 12,118 kilograms of concrete, a 12.2 hour print at 15 millimetre layers and 200 millimetres per second, with ambient temperature carried through the model because it drives cure.

The building itself is parametric CAD written in code and exported to STEP and STL, with a separate module placing electrical, plumbing and HVAC runs against the National Electrical Code and the plumbing and mechanical codes. Structural, permitting and integration gaps were each audited in writing before anything was costed.

Electric hydrofoil: parametric design as code, personal

A board defined entirely in code, with Python generators emitting 37 printable and 32 exchange solids, and design invariants encoded as runnable tests rather than checked by eye.

A parallel adversarial audit by four specialist reviewers found that a claimed structural safety factor of 8.4 had been computed against direct shear, when the real load case is a bending moment. The true figure was 0.22, nine times under-designed. Redesigned to 2.69 while taking the part from 5.6 kilograms to 558 grams.

Also built

  • A nightly retrieval-quality evaluation harness. Investigating a long-running red state established that none of 100 stored runs had ever passed all three gates. The thresholds were stock values never calibrated to this corpus. Recalibrated against a full replay so 67 of 100 pass and every remaining failure is a genuine collapse.
  • An agent fleet and a protocol for letting agents do risky work. 39 specialised agent definitions with an explicit dispatch index, over a written seven-stage protocol carrying 279 tests: specification, review by an evaluator in a separate context, an isolated worktree, and continuous scope-drift monitoring during the build.
  • Tool servers where the risk model is the product. A 35-tool server exposing a customer relationship platform to AI agents in which every call is risk-classified: sending and billing refused outright, writes gated behind a stop sentinel. A public catalogue server built alongside it was hardened across eight adversarial review cycles to 299 passing tests.
  • An internal operations platform. A zero-dependency server behind single sign-on that replaced a commercial work-management product: 25 pages, three real authorization layers that fail closed, and pricing mathematics shared byte-for-byte between browser and server so the two cannot disagree.
  • Bot detection in web analytics. 8,666 of 14,060 reported sessions over thirty days identified as automated by physically impossible viewport resolutions, with a built-in validity check. The business now reports from the cleaned series.
  • Making the catalogue legible to AI buyers. A machine-readable surface trimmed from 467,503 characters of theme chrome down to 6,289, a crawler policy that blocks training bots while admitting retrieval bots, structured data across the catalogue, and 518 of 518 metadata writes applied in 90 seconds with zero failures. Then measured honestly: 81 assistant visits in 30 days across four AI products, with no attributed revenue yet.
  • Finance automation that refuses to guess. A scheduled refresh of the accountant-facing workbook that will not write a single cell until closed months reproduce from the accounting system to the cent. January through July reproduced 33 of 33.

How I work

  • Measure, publish, and retract when the measurement does not hold. Several results on this page exist because I disproved my own earlier claim and said so in writing.
  • Fail closed on anything that reaches a customer, with the fail-open exceptions named and justified one at a time rather than inherited by accident.
  • Deterministic where determinism is possible; a model only where judgment is required. Extraction stays in code and language models judge. A deterministic classifier once recovered 602 records a model had scored as false negatives.
  • I build with agents, and I say so. Most of the volume above was produced with AI coding agents under a protocol I wrote for exactly that purpose. The engineering judgment that matters is what to let them touch, what has to fail closed, and which of their results survive an independent measurement.
  • A written change-safety regime. Three change classes with escalating approval, a named critical-systems list, and a mandatory pre-change callout covering affected files, behaviour delta, hidden workflow risk and rollback path.

What this is not. I build applied AI systems and the production engineering around them. I have not trained foundation models, written CUDA kernels, built distributed training infrastructure, or done machine learning research. If a role needs that, I am the wrong candidate and would rather say so here than in the third interview.

Much of the implementation volume above was produced with AI coding agents, under a protocol I wrote for that purpose. The engineering that matters is deciding what they may touch, what has to fail closed, and which of their results survive an independent measurement.

Experience

Sales Manager, Innovative Bioscience

January 2024 – Present  ·  Las Vegas Valley, Nevada

Sales lead for cell culture reagents, laboratory consumables and nitrile gloves into research universities, cancer centres, biotechnology companies and manufacturing facilities. Also the builder of the commercial systems the sales organisation runs on. Everything above was built in this role.

  • Designed, built and deployed the e-procurement integration that lets institutional buyers purchase inside their existing purchasing platforms, owning both the technical integration and the supplier-enablement conversation with procurement teams on the other side.
  • Classified the product catalogue to the UNSPSC standard and published it through structured data, punchout catalogues and marketplace feeds so procurement platforms can read it.
  • Consolidated customer relationship management, quoting, price sheets, order history and accounting into a single reporting layer, replacing five systems that disagreed with one another.
  • Built the quoting and price sheet engines the team uses daily, including institutional pricing tiers, an approval gate below a costed floor, and tax-exempt handling.
  • Rebuilt the commerce experience around how technical buyers actually buy: contract pricing, quote-first conversion and technical depth rather than impulse checkout.
  • Ran a multi-state sales tax exposure review across the full customer book and drove the exemption-certificate remediation that came out of it.
  • Wrote the sales training programme and onboarded new territory representatives end to end, from systems access to account handoff.
  • Administered the Microsoft 365 tenant and identity estate, including a gap sweep that found an outgoing vendor still holding three enabled identities and two paid seats after a remediation reported complete.

President, SunNSee.com

June 2019 – November 2023  ·  United States

Founded and ran an e-commerce and equipment business serving ophthalmic and optometric practices nationwide: sourcing, refurbishing and selling diagnostic and surgical equipment, and building the storefront and fulfilment operation behind it. Started while still carrying the Lombart territory and went full time in 2020.

  • Launched and ran the e-commerce platform for ophthalmic and surgical equipment, owning the storefront, merchandising and the customer relationship end to end.
  • Owned supplier relationships and nationwide fulfilment.
  • Produced every piece of marketing content the business ran, including the product photography, the video and the aerial drone work. No agency, no contractor.
  • Refurbished and resold Briot and Weco optical edgers into private practices, running the technical demonstration, installation and integration training myself.
  • Advised practices on return-on-investment equipment upgrades and lane layout. It is the same consultative motion I now run into research institutions.
  • Built an inventory tracking system to streamline business-to-business fulfilment. It was the first commercial software I wrote against my own profit and loss, and the reason I now build the systems rather than buy them.

Regional Sales Manager, Lombart Instrument (Advancing Eyecare)

October 2018 – August 2020  ·  Nevada and Utah

Carried a regional territory for a national ophthalmic equipment distributor, selling capital diagnostic and surgical systems into private practices, surgical centres and hospital systems. Long-cycle, technical and consultative. It is the same sales motion I run now, into a different building.

  • Managed a $1.8 million annual sales pipeline across diagnostic and surgical ophthalmic systems, including optical coherence tomography imaging, corneal topographers, YAG lasers and auto-refractors.
  • Delivered operating-room and clinic product demonstrations, technical presentations and multi-department staff training.
  • Ranked in the top tier of the sales force through a consultative, clinically aligned approach, and named Top Regional Sales Performer in 2019.
  • Partnered with hospitals and private practices through installation and onboarding, so the sale did not end at the purchase order.
  • Ran the capital-equipment buying process end to end: return-on-investment modelling, leasing programmes, and multi-stakeholder approval across surgeons, operating-room staff, administrators and purchasing. That is the same committee I sell to today.

Ophthalmic Technician, private practice

2014 – 2017  ·  Nevada

Clinical side of the same equipment I later sold. Performed pre-operative and post-operative testing for LASIK and cataract procedures, calibrated the diagnostic instruments, and counselled patients through their results.

This is why an ophthalmic equipment distributor hired me: I had already run the devices in a clinic before I ever demonstrated one to a surgeon.

Optician, Eyes on Las Vegas

2011 – 2015  ·  Las Vegas, Nevada

Dispensing optician. Fitted and adjusted eyewear, took the measurements, and cut and edged lenses to prescription. Worked the bench while completing the ophthalmic laboratory technology degree, and stayed on it as the clinical technician work began.

This is where the optical thread starts. I was making the lenses before I ran the diagnostic equipment, then sold the equipment, then built the software.

Independent and client work

Pharmaceutical manufacturing data platform, client engagement 2026

Leading the digital side of a data platform for regulated pharmaceutical batch review: message architecture, site, and an interactive product demonstration built on a public production dataset of 1,005 batches and 4.7 million process samples. Vocabulary grounded in 73 sections of federal regulation pulled from the government's own interface, with a competitive capability matrix built across twelve competing products from their own documentation.

Two findings mattered more than the build. Charted series were being drawn as window means, understating one true peak by 80 percent. On a product whose entire subject is data integrity, a chart showing a smoothed derivative of the record is the one defect you cannot ship. Replaced with true minimum and maximum envelopes.

And the analysis contradicted the expected story. Process excursions did not predict the quality outcome at any threshold, while material genealogy did: one raw material lot produced six out-of-specification batches in twelve, against a base rate of 7.4 percent.

Technical

Design & media

AutoCAD  ·  parametric CAD in code (CadQuery, STEP and STL)  ·  Blender headless rendering  ·  generative image and video pipelines with automated quality gates  ·  ffmpeg  ·  Remotion  ·  scripted browser capture  ·  drone aerial capture  ·  brand and marketing content production

Clinical & capital equipment

Optical coherence tomography  ·  corneal topography  ·  YAG lasers  ·  auto-refractors  ·  Briot and Weco optical edgers  ·  RTSP and IP-enabled operating-room camera systems  ·  capital equipment return-on-investment modelling and leasing  ·  technical installation and clinical staff training

Practice

Test-driven development  ·  adversarial and red-team review  ·  measurement and evaluation design  ·  fail-closed safety architecture  ·  technical writing and system documentation

Education

College of Southern Nevada Associate of Science, Ophthalmic Laboratory Technology 2011 – 2013
Front Range Community College Applied Science, Mechanical Engineering 2009 – 2011