Things I've built. Most of them run a dental practice every day.

Problem

A 10-year-old single-price tool couldn't handle three clinics with different price books. The 6-month price-lock was a promise on the PDF, not a system property. An earlier vendor attempt never shipped.

Solution

Next.js 16 + Postgres + Drizzle rebuild. Five pricing models (flat, tiered, tiered-zoned, price-range, per-surface). Captured `*_at_capture` columns freeze every saved price, with per-location price book versions.

Process

9-decision ADR with tradeoff tables and library citations before any code. 100% branch coverage on the pricing engine. Golden snapshots lock every documented wizard path. A `questions-for-Chelsea` doc with worked numbers lets the clinical SME review without reading code.

Result

Shipped end-to-end in about a month, what a vendor failed to ship in a decade. Runs under $35/mo. Since then, 72 tagged releases through October 2026: a three pane drag and drop workspace with phases and scenarios, an Excel export that matches the coordinator's own file with live formulas, a landing page that is the day's consult schedule, visit sequences built from the estimate, and a full restyle in the Hybridge design in nine releases on one day.

Problem

Coordinators wrote each patient's visit plan by hand in Word, copied from master documents in Drive, in two formats: a flat list of chair time units, and a six column table for full arch cases and dentures. Every change meant renumbering the visits and redating everything after it by hand, on a case that can run nine months.

Solution

Two packages. sequence-core holds the domain model, the practice masters stored as data, generation from an estimate, date and wait arithmetic, and the Word and text renderers, with no React. sequence-ui is a three pane React editor: visit list, flowchart and a live document preview. The estimator stores a living plan plus immutable version snapshots, and one shared text module keeps the preview, Word and PDF from drifting apart.

Process

Generation picks a template; it never infers a plan. A procedure with no matching visit prints as a yellow open item, an unsupported combination asks for a manual build, and a free text wait stops date projection instead of guessing. Anything not transcribed from a master carries a DRAFT banner on screen, in the PDF and in Word. The packages reach the estimator as a committed build snapshot stamped with its source commit, and a check fails if core and ui come from different commits.

Result

Live inside the estimator since 27 August 2026, with save, versions and PDF a release later, a read only lookup for the patient dashboard, and a Word export that now embeds Inter instead of opening in Times New Roman. A dev database audit matched 84 of 92 generation attempts to a validated master with 0 errors. The named tradeoff: nothing fails yet when the two repos fall out of step, so automating the refresh is next.

Problem

The April version on Apps Script replaced a live API dashboard that broke constantly, but it claimed more than it delivered. Every metric had three implementations (card, drilldown and trend overlay), so two tabs could disagree. The connect column holds only the surname, so about 31% of patients resolved to an arbitrary lead: the 49% to 99% jump measured coverage, not correctness. A multi year Weekly view took about 5.2 seconds.

Solution

A rebuild on Cloud Run behind IAP. Cloud Scheduler checks Drive every five minutes; a private ingest service parses the four exports with ExcelJS, joins and hashes them, and writes an immutable snapshot to Cloud Storage. The browser loads one pinned snapshot of about 0.4 MB and computes everything locally with the same core package the ingest uses, and each card number is a function of its drilldown rows by construction.

Process

Business rules were ported verbatim and held to a parity gate that runs the original Apps Script code in a Node VM. When parity could not catch a one day date shift (both sides used the same reader), a golden test pinned figures read off the old production dashboard. A sorted date index means a one week query touches about 60 rows instead of scanning every lead, and a request sequence guard stops stale results from landing.

Result

Live on Cloud Run since 2 September 2026. The Weekly tab dropped from about 5.2 seconds to about 17 ms, ambiguous links fell from 31% to 5.8% with link quality recorded on every row, a Cash Flow tab shipped on 15 September, and on 30 September it took on the Hybridge design while staying dark. The weekly routine did not change: drop four files.

Problem

A three person insurance team spent about 70 hours a week cross checking the schedule, the production ledger, and the clinical notes by hand to confirm every patient seen got documented and billed under the right doctor. It was slow and it failed exactly where mistakes cost the most: care documented but never billed, billed with no note behind it, or billed under the wrong provider.

Solution

A pure, deterministic reconciliation engine wrapped in a thin FastAPI app. Two source adapters (Eaglesoft CSV, Denticon xlsx) normalize hostile exports into the same typed models, the matcher keys everything on patient_id, and a stack of checks (money vs note, billing provider vs the doctor named in the note, note-claimed codes vs the ledger, perio charting proof, office policy rules) separates verified patients from the handful that need a human.

Process

The engine has no network, no database, and no clock, and it takes the timestamp as a parameter, so identical inputs give byte-identical output and every count traces to a source row. Columns map by name and a missing one fails loud; the app also checks the dates inside every file and warns when an export covers the wrong day. 200+ tests run on synthetic fixtures that mirror every real quirk, plus an opt-in golden test against a real day, so no PHI ever enters the repo. Every false flag the team reported got traced, fixed, tested, and redeployed the same day.

Result

Shipped and live on Cloud Run, in daily use. About 70 hours a week of manual cross checking became an exception review of roughly a dozen flags out of ~120 patients a day, over 3,600 hours a year returned. A Gemini clinical matcher is built and waiting behind a preview-mode rollout.

Problem

The VP of Finance had built a sound forecast in Excel and handed it over with a spec, but it needed a manual paste, recalculate and send every morning, fed by two daily export emails from the dental lab's case system.

Solution

A pure Python engine package with no I/O, and a pipeline package with parsers, a dedupe ledger and a BigQuery loader, with Terraform for the infrastructure. A daily Cloud Run job runs the model, and a Cloud Run web app behind IAP shows the dashboard, history and a Projections tab. An Apps Script bridge runs inside the VP's own Google account and copies the report attachments to a landing bucket, so the project never holds a credential that can read his mailbox. Gemini writes the commentary over finished numbers only.

Process

Parity first: every acceptance value and edge case from the spec is pinned on every commit. Every number in the Gemini commentary must appear in the report payload, a list of banned phrases blocks advice the model has no business giving, and if the guard fails the report falls back to a deterministic template. No auto tuning: the best in sample parameter grid scored 2.59% but was 0.62 points worse out of sample. A better lag model sits behind a switch because its 0.28 point gain did not clear the noise floor.

Result

In production since 25 September 2026. Current month revenue error falls from 4.2% early in the month to 1.5% late, and the report never depends on the LLM. The one hand typed figure now lives in BigQuery as rows that supersede instead of overwrite, after a scale to zero quietly lost it while the form still said Saved.

Problem

Marketing numbers lived in five separate logins, and no off the shelf connector could join ad spend all the way through to treatment value, so a connector product was ruled out.

Solution

Google Ads lands through the BigQuery Data Transfer Service; Microsoft Ads, Meta, CallRail, Monday CRM and TapClicks come in through Python connectors running as Cloud Run jobs on Cloud Scheduler. Data moves from raw to staging to marts, every column carries a description, and a data guide tells an LLM what to read first. People query it through Looker Studio, Gemini and an MCP connection.

Process

A daily ops job checks freshness and runs a drift check that alerts on a z score above 3 against a 28 day baseline, and a weekly eval harness asks Gemini 22 golden questions. The lead to patient join moved from fuzzy phone, click id and email matching to the CRM's own board relation link, with the match method recorded on every row. View types that carry the identities of callers, reviewers or leads are blocked in code, whatever the config asks for.

Result

Live since early August 2026, with the TapClicks connector added on 25 September. 98.5% of patients link to their lead, and the join agrees with the weekly exports 99.5% of the time. It is not perfect: the first eval run passed 77% against an 80% gate, and on 23 September pipeline health stood at 32 of 36 checks with the CallRail sync failing two nights running.

Problem

Per-seat SaaS automations were expensive and getting more so. C-suite analytics relied on manual exports each week.

Solution

Self-hosted n8n on AWS via Docker. One platform for every team automation, with data flows standardized across the org.

Process

Migrated existing automations into n8n. Built executive analytics pipelines that publish on a schedule instead of on-demand.

Result

20% lower automation cost. 500+ hours/year recovered on executive analytics. AI initiatives ship on a shared substrate.

Problem

Doctors write the prescription in the lab vendor's customer portal, which cannot edit an existing case, so every prescription arrives as a second case that duplicates the scan case already in the system, and customer service has to catch each pair to merge it.

Solution

A Next.js app in the Hybridge design on Cloud Run behind IAP. A Cloud Scheduler job runs at 06:00 ET and lists the likely duplicates. The data sits behind a case source and store interface: sample data today, the vendor's API and Firestore once the license lands.

Process

The name matcher forgives case, spacing, a middle initial, swapped first and last names, and a single mistyped letter. Anything more is a different person.

Result

Demo deployed on 30 September 2026 on sample data, with 67 tests. Live mode is blocked on a vendor API license, and a second phase doctor portal is parked.

Problem

Website forms on four company sites ran through a chain of three vendors: Tally to collect, Zapier to move, Monday to land.

Solution

One owned platform: a public forms app with an embed script, an admin dashboard with a form builder behind IAP, a read API for downstream tools, and nightly checks.

Process

Built around one rule: no submission is ever lost. If Firestore fails, the answers go to a private Cloud Storage bucket and the visitor still sees success. If both fail, the server answers 503 and the browser retries four times over about 15 seconds with the same id, then keeps the answers on the device. An importer brings over all 32 live Tally forms.

Result

In build: about 11 of 18 build tasks done as of October 2026, with over 1,300 tests across the core, admin, renderer and forms packages. Not deployed yet.

Problem

"Detect fake news" hides two problems: labeling whether a claim is checkably false, and reading how a crowd of 1.8M tweets feels about it. They need different models and different data.

Solution

Llama 3 8B fine-tuned with QLoRA on the LIAR benchmark for misinformation, RoBERTa fine-tuned for sentiment over the tweet firehose, and LDA for topic clusters, all hanging off one shared cleaning module.

Process

Watched per-class precision/recall instead of trusting raw accuracy on imbalanced LIAR. Unified the tweet and LIAR cleaners to kill train/serve skew. Tuned LDA by coherence after stopword/hashtag handling.

Result

76% on LIAR (honest, given short politically-loaded claims), cheap RoBERTa inference across 1.8M rows, and defensible topic clusters. Route the cheap model over everything, spend the 8B only on flagged claims.

Problem

A deep-learning class competition: 55 teams, two weeks, best CIFAR-10 classifier. Hard rules: stay a ResNet, stay under 5M parameters. We cleared the 85% bar and then sat at a frustrating 42nd.

Solution

The rules said no bigger model, so I added data instead. Augmentation, rotating, cropping, and flipping the CIFAR-10 images, gave the model far more to learn from without adding a single parameter.

Process

I tried different ResNet variants and swept hyperparameters to a solid setup (~92%, 16th in class), but tuning hit a wall under the cap. I read papers and open-source repos to see how the best were doing it, then reframed the problem from 'better model' to 'more data.'

Result

97.12% accuracy, 2nd of 55 teams (0.17% behind 1st). On the professor's secret unseen test set, our model generalized best and finished 1st in the whole class. Earned a bonus grade point.

Problem

At 22M+ records on HDFS, scoring every user against every item is intractable. The question stops being "which model" and becomes "how do I not compute the things I don't need to."

Solution

A two-stage pipeline: MinHash LSH buckets similar entities for cheap, recall-oriented candidate generation, then ALS matrix factorization ranks only the few hundred survivors per user.

Process

Salted hot keys and repartitioned to fix Spark data skew. Tuned LSH bands/rows against candidate recall as an explicit knob. Fell back to the popularity baseline for cold-start users.

Result

+20% Precision@K over the popularity baseline (the honest thing to beat) at 22M+ records. Swap LSH for ANN and ALS for a learned ranker and it's the literal production blueprint.