Tyler Brown

Field sales platform for a residential solar contractor

Forward-deployed with the client's VP of Sales and field reps since September 2026. One Progressive Web App the sales team works from in the field: door-knocking maps and territories, leads, rep tracking and team analytics, a roof designer, proposals, training, and onboarding with e-signed contracts. Serverless REST API on Cloudflare Workers, D1 (SQLite) and R2, behind single sign-on, with GitHub Actions CI on every push.

The roof designer takes Google Solar API roof segments, then fits roof planes to public USGS LiDAR point clouds in the browser to cross-check and correct them. When reps said layouts were too strict, I replayed real homes instead of arguing: rows matched LiDAR, the strictness was real, and placed panels rose from 608 to 626 with all 948 tests passing.

Delivery runs from the field: reps file tickets in the app, I turn each into a fix with Claude Code, Codex reviews it adversarially, and a release gate tests the exact commit on a frozen copy, ships to staging first and smoke-tests production. The gate refuses any build it cannot prove.

Built and awaiting release: Pitch Coach, an AI "skeptical homeowner" (Llama 3.1 8B on Workers AI) that reps practise door pitches against, scored on a five-point rubric with per-user and company spend caps that fail closed.

Literature extraction on local GPU inference

A pipeline that reads the Methods section of published open-access papers, extracts structured facts, here which commercial supplier each laboratory used, with every answer tied to a verbatim quote that code checks against the source. 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 on the runs that followed, including one unattended 6,537-paper job.

I then distilled the large model into a small one: a 421-million-parameter ModernBERT classifier fine-tuned with PyTorch and Hugging Face on 23,984 labels the large model produced, on the same consumer GPU. Zero-shot it scored 0.47; fine-tuned it scored 0.985 on 200 held-out papers, beat a TF-IDF baseline of 0.86, and runs 68 times faster.

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.

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. Liveness and dependency health are checked separately.

Nine rounds of adversarial security review closed more than forty findings, including injection and session-handling findings, all fixed.

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 within seconds, 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 live against the carrier. 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.

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, searches for the best track position, and emits G-code with pump control. From the best single position the arm reaches half the footprint, so the plan calls for repositioning.

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.

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.

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.