14 projects, ranked by technical depth. Select one to see the business problem, the result and how it was built.
Includes nine builds from the 2026 Gauntlet AI Fellowship, an intensive AI engineering program in Austin, TX.

Postpartum hypertension can cause strokes while dangerous home readings wait in a nurse’s queue. Nurses lose time piecing together charts, delaying action and exposing patients to harm and hospitals to readmissions and liability.
Reliable warning-sign handling and cited chart answers help nurses assess urgency; passing tests with recorded answers does not establish faster clinical action or fewer readmissions.
Code ranks the blood-pressure queue using fixed rules and refuses to lower a rule-assigned priority. The supervisor model explains priorities and investigates charts through tools with validated inputs and patient-access checks; every fact needs a citation. Week 2 adds workers for document extraction and guideline retrieval, while code verifies facts and blocks conflicting priority explanations.

AI products lose context when user information lacks a shared structure for relationships and uncertainty. Separate records leave agents without a consistent foundation for reasoning and personalization.
Preserving sources and uncertainty gives agents a shared foundation for reasoning and personalization without losing where information came from.
A schema-driven graph represents people, memories and relationships in Neo4j. Pydantic validates record structure, field-origin tracking preserves where information comes from, and approximate time representations retain uncertainty instead of forcing exact dates. LangGraph supports coordination between agents using that shared structure.

Importers lose margin and time when wrong tariff codes cause overpayment, delays or compliance exposure. Manual research struggles to keep pace with changing tariff lines, while generic automation can answer confidently without relevant rulings.
Correct codes reduce classification errors: the tuned model improved from 42.0% to 59.8%, gaining 17.8 points, but missed the larger model’s 71.6%.
QLoRA, a low-cost fine-tuning method, teaches a 4-billion-parameter model to choose a tariff code using five retrieved rulings and 20 schedule lines. Code retrieves evidence and checks exact answers. A fixed 500-item test gives the original, tuned and larger models identical evidence for a fair comparison.

Clinical AI teams risk patient harm and reputational loss when security testing stops at static checklists. One-shot prompts miss attacks that adapt to defenses, while weak reporting can mistake outages or uncertain judgments for vulnerabilities.
The probes found no potential breaches, but their scoring method could not directly confirm a breach; the required three confirmed vulnerability reports remained unmet.
Code selects tests within budget, an attack agent adapts prompts to observed defenses, and a shared runner captures live responses. Code checks and one bounded model review assess each attempt, then guide the next attack. A duplicate detector stops repeated prompts, and report generation requires a confirmed finding saved in the evidence store.

Farmers lose productive time when machine performance problems go unnoticed. Without consistent baselines across the fleet, separate telemetry reports make it difficult to spot deviations early and prioritize diagnosis.
Faster anomaly detection helps diagnostic teams investigate machine performance problems sooner and protect farmer productivity.
Automated Databricks pipelines use PySpark and SQL to aggregate machine telemetry and establish consistent performance baselines. Power BI dashboards compare incoming performance against those baselines and surface deviations through real-time alerts, giving diagnostic teams a shared view of emerging problems.

Home-service sales teams waste paid field time walking streets without evidence about which doors fit a campaign. Manual property research leaves gaps, while unexplained scores give representatives little reason to trust a suggested route.
Complete coverage supports territory planning: all 553 homes were searched in 7 minutes 48 seconds, although 3 image results remained unassessed.
A fixed seven-stage workflow gathers property evidence. An image model detects features, and a second model checks pools crossing property lines. Code controls eligibility, scoring, wording and routes; invalid answers become “not assessed.” Every call reserves budget, every score retains its source and date, and labeled development cases plus five previously unseen homes test image quality.

People lose useful insights when their personal information stays scattered across platforms. Separate applications lack a shared picture of the person, limiting meaningful personalization.
A shared view of personal information gives AI workflows context for meaningful insights and personalization.
The application combines personal data with facial recognition and AI-generated insights to support personalization. LangChain organizes model workflows, while application code manages authentication, file storage and domain-specific services. A modular backend separates those responsibilities.

Relationship-support products lose trust when remembered claims lose their source, merge the wrong people or survive deletion requests. Saving model summaries directly can turn plausible text into false memories without source evidence, confidence or consent controls.
Faithfulness measures whether extracted memories stay grounded in their source: 0.94 improved on the original model’s 0.50, but final testing on reserved examples remained outstanding.
I co-own Memory with a teammate. Models propose claims, while Pydantic validation requires source evidence and confidence. A Python engine controls the Postgres work queue, source-quote checks, identity matching, memory linking, access restrictions, storage and erasure. Wording checks reject unsafe terms or checking failures, spending limits bound calls, and a fixed 400-row set awaits final testing.

Prediction-market intermediaries risk misrouting capital when similar-looking contracts settle under different rules. Comparing titles and prices cannot establish equivalence because thresholds, evidence sources and cancellation policies may differ.
False matches can send capital into contracts with different payouts; none occurred, but 505 candidates for 34 planted pairs triggered a required discovery redesign.
Models extract ambiguous contract terms, preserving source passages and the model, prompt and output-format versions. Code compares seven dimensions through two equivalence checks, applies a third routing check, and simulates filling the entire order or none. Missing or conflicting terms stop the contract from proceeding, and saved inputs reproduce decisions exactly.

Home-services marketplaces lose time and trust when provider histories, schedules and availability disagree across systems. Combining spreadsheets does not reliably identify duplicate providers, preserve where records came from or protect homeowner details.
Protecting homeowner data preserves trust: no answer contained detected personal data, although summary counts incorrectly flagged ordinary uses of a word that also appears as a surname.
Seven read-only tools let an AI assistant query one shared database. Separate importers normalize each source, while code resolves provider identities and records each match’s confidence and origin. An output scanner masks personal data, and a 58-question evaluation checks tool choice, honesty and privacy using development questions and previously unseen variations.

Care providers lose enquiries and acquisition spend when anxious families repeat their story, retype identity details and abandon booking. A generic chat form cannot safely interpret free text, handle identity documents or make commitments without confirmation.
Accurate identity details reduce rework and preserve trust: name, birth date and address were correct on 30/30 cards, including correctly leaving absent fields blank, with no fabricated values.
One model call per chat turn interprets intent and drafts wording. Local text recognition reads IDs, while code controls accepted fields, dates, consent, booking and database writes. Before each model call, a filter replaces recognized identifiers; invalid model answers trigger a safe fallback, and every ID field requires confirmation before storage.

Roofing operations lose hours checking permit requirements for each ZIP code. Repeating zoning searches manually creates a bottleneck as coverage expands.
Shorter lookup times remove a recurring operational bottleneck and make broader ZIP-code coverage practical.
A modular Python pipeline uses Playwright to collect zoning information and GeoPandas to organize location-based results. Parallel schedulers on Google Cloud divide the lookup work and avoid duplicate processing, allowing the same collection and aggregation process to cover additional ZIP codes.

Contractors lose sales and lenders lose applications when homeowners stall at product selection or income verification. A general chatbot adds compliance risk because it can invent payments, imply approval or expose sensitive application data.
Blocking every crafted violation demonstrates prototype safeguards against compliance and trust failures; testing against a reviewed set of real production answers remained future work.
Models explain product terms, extract estimates and prepare applications using four tools. Code calculates figures and disclosures, assembles fields and intercepts approval questions. Checks reject uncomputed figures or approval language, retry once, then show a safe fallback. An append-only log records every attempt; limits stop work at six steps, 45 seconds, 40,000 tokens or three repeated calls.

Founders lose weeks researching startup opportunities before they can test an idea. Manually reviewing reports and market trends limits how quickly they can compare problems across industries.
Reducing research effort lets founders assess opportunities sooner and compare business problems across industries.
Continuous n8n workflows send consulting insights, industry reports and market trends to models through the OpenAI API. Prompts extract business pain points and synthesize research into opportunity concepts, while workflow automation coordinates repeated processing.