# Greg Schwartz, AI Engineer > AI engineer building evals, RAG and LLM features. Stanford HCI master's, host of The Prompt Engineering Podcast, US patent holder. For teams hiring an AI or evals engineer. Stanford MS in HCI. 21 years in software, the last 4 in AI. - HIPAA RAG support chatbot: patients grew 6x, med team only 1.9x because it resolves 67% of patient messages - Eval'ed prompt chain, 6x speedup for doctors writing reports - Support triage accuracy 57% to 96% with golden sets + evals - Lab PDFs to EMR 2.8x faster, 95% field accuracy Host of The Prompt Engineering Podcast, 87k downloads. US patent 10049598. - Brand: Greg Schwartz, Gregory Schwartz, The Prompt Engineering Podcast --- # Before AI: two redesigns that closed funding Source: https://gregschwartz.net/before-ai.md [← work](/#work) Before AI · product and UX, 2014 to 2021 # Two redesigns that closed funding Before I shipped AI, I redesigned products their users found too confusing. These two are the ones that moved the business. [ Sighten.io · solar software $3.5M Series A after the redesign Users and investors said the app was too confusing. I redesigned it top to bottom, starting with a navigation bar that hid the workflow. case study → ](/before-ai/user-experience-design-interaction-design-wireframing-massively-streamlined-navigation) IPVM 12 a day Invites after the redesign Up from one invite every two days. [next → public work](/public-work) [contact](/contact) --- # Massively Streamlined Navigation Which Got Funding Source: https://gregschwartz.net/before-ai/user-experience-design-interaction-design-wireframing-massively-streamlined-navigation.md ![Final version](https://gregschwartz.net/wp-content/uploads/2016/02/12-final-version-1024x720.png) ### Project Description Users and venture capitalists had said the client's web app was too confusing. VCs were refusing to invest for that reason, too. I met them at a networking event and they asked if I could come in the next day. Amongst many issues, the application had a "summary bar", which both displayed content **and** was used for navigation. It wasn't clear that it could be clicked on, and it wasn't clear that it was meant to indicate a workflow, rather than distinct pages. After my top-to-bottom redesign, they quickly were able to raise funding! ### Skills Used - Interaction design- Wireframing ### Before I came on [![1 before](https://gregschwartz.net/wp-content/uploads/2016/02/1-before-1024x681.png)](https://gregschwartz.net/wp-content/uploads/2016/02/1-before.png) They had already received strong user feedback that the blue-green "summary" bar was confusing. It wasn't obvious that you could click it -- there's nothing "clickable" about it! Worse, it didn't indicate that there was a workflow (e.g. you had to create a "Site" before creating a "System"). [![2 before](https://gregschwartz.net/wp-content/uploads/2016/02/2-before-1024x889.png)](https://gregschwartz.net/wp-content/uploads/2016/02/2-before.png) When it was completely filled with data, it was almost overwhelming. ### Wireframes To show it was a workflow, I turned the bar into arrows. Then I played with various bar and pie charts to visually show the percentage of energy generated from solar. (It's faster to interpret a ratio in a simple graph, than with text labels.) #### Variations ![](https://gregschwartz.net/wp-content/uploads/2016/02/3-wireframe-two-charts.png)![](https://gregschwartz.net/wp-content/uploads/2016/02/4-wireframe-simpler-lables.png)![](https://gregschwartz.net/wp-content/uploads/2016/02/5-wireframe-attach-bar-to-top-nav.png) Multiple layers of navigation is confusing, so I tried merging the summary bar with the top navigation bar: [![](https://gregschwartz.net/wp-content/uploads/2016/02/6-wireframe-merge-nav.png)](https://gregschwartz.net/wp-content/uploads/2016/02/6-wireframe-merge-nav.png) [![](https://gregschwartz.net/wp-content/uploads/2016/02/7-wireframe-simplify-merged.png)](https://gregschwartz.net/wp-content/uploads/2016/02/7-wireframe-simplify-merged.png) [![](https://gregschwartz.net/wp-content/uploads/2016/02/8-wireframe-simplify-more.png)](https://gregschwartz.net/wp-content/uploads/2016/02/8-wireframe-simplify-more.png) Ultimately, they decided they wanted the navigation separate, so I attached the "summary bar" to the top navigation, to make it clearer that it was still navigation. [![9 wireframe - separate from nav](https://gregschwartz.net/wp-content/uploads/2016/02/9-wireframe-separate-from-nav.png)](https://gregschwartz.net/wp-content/uploads/2016/02/9-wireframe-separate-from-nav.png) ### So much better! Once their engineers implemented it, it significantly improved the readability of the page. It also made it much easier to focus on the purpose of the page, instead of the huge summary bar at the top of the screen. [![](https://gregschwartz.net/wp-content/uploads/2016/02/10-implemented-early-in-process.png)](https://gregschwartz.net/wp-content/uploads/2016/02/10-implemented-early-in-process.png) [![](https://gregschwartz.net/wp-content/uploads/2016/02/11-implemented-later-in-process.png)](https://gregschwartz.net/wp-content/uploads/2016/02/11-implemented-later-in-process.png) ### Further improved into final state Optimally, navigation and a summary should be separate, but they had wanted to keep them together. A front-end engineer and I worked together to prototype a further simplification, with summary on the left (reduced to just name and address!) and navigation on the right. [![12 final version](https://gregschwartz.net/wp-content/uploads/2016/02/12-final-version-1024x720.png)](https://gregschwartz.net/wp-content/uploads/2016/02/12-final-version.png) --- # Public work Source: https://gregschwartz.net/public-work.md # Public Work Now playing ## The Prompt Engineering Podcast Host and producer since 2023. Teaching prompt engineering to engineers and laypeople. [Listen](https://PromptEngineeringPodcast.com) 0downloads 0downloads an episode, about 0episodes 0countries 43% United States 31% Europe 26% other Patent ## Passive tracking of food via wearable body sensors Filed while Director of Product at Zipongo; the idea came out of the BodyMedia talks below. US10049598B1 Paper · UIST 2008 ## Harvesting Helpfulness Studying an Online Farmer's Forum. A/B tested enhancing forum threads with maps of commenters' locations. ## Talks [30:33Giving Good Presentations: stop putting paragraphs on your slidesWatch](https://www.youtube.com/watch?v=RIq63685J2Q) [7:56Quantified Dating: data analysis of online dating messages, photos, and profilesWatch](https://www.youtube.com/watch?v=ZVYxZj4WIiY) [5:07Burning BodyMedia: physical activity and calorie expenditure at Burning ManWatch](https://www.youtube.com/watch?v=x-glCz7zVBk) [5:20Are You What You Eat, or What You Buy? Caloric intake, activity, bodyweight, and nutrients · became the patentWatch](https://www.youtube.com/watch?v=B1k2qosppCk) ## Awards 2011iOSDevCamp, Best Audio/Video App · VJMixer 2008iPhoneDevCamp, Best Game · Tattletalz 2005UCI Business Plan Competition, Best Undergrad Team and Third Place · an eBook reader and content delivery system, two years before the Kindle 2023Founded San Diego's first AI hackathon · 180+ attendees, 18 sponsors ## Community 2016 to 2020Organizer, Quantified Self San Francisco mentorSpringboard · 96.2% mentee satisfaction 1999Eagle Scout [off clockFire dancer, poiPhotos](https://gregschwartz.net/fire-dancing) --- # 150+ prompts for GoDaddy Source: https://gregschwartz.net/godaddy-prompt-library.md [← work](/#work) GoDaddy · prompt engineering · evaluation # 150+ promptsfor GoDaddy I wrote, rewrote and tested over 150 prompts for GoDaddy's free AI prompt library for small business owners. readout153 prompts rewritten111 written new42 gpt-4-turboOpenAI APIRubySQLiteGoogle Sheets ## The problem GoDaddy was building a free prompt library for small business owners. Many of the first drafts gave generic output, nudged the model to make things up, or asked for work no model could do yet, like live web research. ## How I tested Every prompt ran against three test businesses: a burger restaurant, an online jewelry store and an auto detail shop. A Ruby script I wrote sent them to gpt-4-turbo three at a time and cached each result by a hash of the prompt, so the whole set could rerun after every change. I read every output and fixed what broke. 56 of the first 81 prompts changed. ## By category Industry prompts (6 industries): 30SEO: 13Content creation: 11Digital marketing: 11Sales: 11Business operations: 10Customer support: 10Social media: 10Ad campaigns: 8Branding: 7Customer loyalty: 7Data analysis: 7PR and outreach: 7Starting a business: 6Productivity: 530131111111010108777765153prompts Industry prompts (6 industries)30 SEO13 Content creation11 Digital marketing11 Sales11 Business operations10 Customer support10 Social media10 Ad campaigns8 Branding7 Customer loyalty7 Data analysis7 PR and outreach7 Starting a business6 Productivity5 ## Live on GoDaddy [![GoDaddy](https://gregschwartz.net/wp-content/uploads/godaddy-ai-prompt-library.webp)](https://www.godaddy.com/resources/ai-prompts-for-small-business) GoDaddy's AI prompt library for small businesses [![GoDaddy](https://gregschwartz.net/wp-content/uploads/godaddy-quick-start-prompt.webp)](https://www.godaddy.com/resources/ai-prompts-for-ad-campaigns) One of my prompts, live on GoDaddy's ad campaigns page > What the client said > Greg completed the most incredible prompt I've ever seen, having spent a ton of time myself researching it and not finding any examples anywhere. Greg was very aware of the needs of our audience, adding elements to his prompts to ensure output was optimized, reducing the fluff that tends to come with Chat GPT. > > > Alycia Leno · Marketing Director · GoDaddy [next → before AI, two redesigns that closed funding](/before-ai) [contact](/contact) --- # Greg Schwartz, AI engineer Source: https://gregschwartz.net/index.md ## Greg Schwartz # AI engineer ### Ship the eval, then the feature. Twenty one years of software, the last four shipping production AI. Trained in human computer interaction at Stanford, so the systems get adopted, not just demoed. See the work [Resume](/resume) [ Doctors' reports 6x faster, from a three step prompt chain ](/doctors-reports) [ Triage accuracy 57→96% ](/support-triage) Podcast downloads 0 19 episodes, 155 countries [ 6x the patients 6x 1.9x 0x the team. Not 6x. ](/member-messaging-assistant) [ Patent Passive food tracking via wearable sensors US10049598B1 ](/public-work) Education Stanford Engineering Master's degree, Human Computer Interaction OpenAI GPT-6 · CodexAnthropic Claude · Claude CodeGoogle GeminiLangfuseBraintrustPostgres RAGFunction calling · MCPElevenLabs · Retell · BlandPython · DjangoTypeScript · ReactHIPAA OpenAI GPT-6 · CodexAnthropic Claude · Claude CodeGoogle GeminiLangfuseBraintrustPostgres RAGFunction calling · MCPElevenLabs · Retell · BlandPython · DjangoTypeScript · ReactHIPAA ## Work result · models · tools · process [ Evals36k tickets 57% → 96% Support triage, brought back from dead The old classifier was a coin flip, so three full time staff plus department heads read every ticket. Golden sets of 50 and 200 human labeled tickets, 15 categories, prompt and few shot iteration scored in Langfuse, then image categorization for scales and prescriptions. tokens per ticket after slimming the input-25.9% ](/support-triage) [ Chatbot and RAG 6x 1.9x expected team actual team 6x the patients. Not 6x the team. Member messaging assistant HIPAA compliant, on GPT-4.1, GPT-5 mini and GPT-5.4 with Postgres RAG over 100+ clinical protocols. Resolves two thirds of messages through an 8x volume increase. The care team should have grown 6x. It grew 1.9x. ](/member-messaging-assistant) [ Structured output 2.8x Faster inputting Labs to EMR GPT-5.5 parses lab PDFs and faxes into structured fields, 95% accuracy on 40 fields, clinician validation view. 9.1 to 3.2 minutes. ](/lab-results) [ MedowHealth.ai · prompt chain 6x Doctors' reports, faster Owned and evaluated a three step prompt chain that turns medical transcripts into finished reports. ](/doctors-reports) [ GoDaddy · prompt engineering 150+ prompts A prompt library for marketing Ad campaigns, branding, content, SEO, sales. "Our go to prompt engineer," per their marketing director. ](/godaddy-prompt-library) Now playing ## The Prompt Engineering Podcast Host and producer since 2023. Teaching prompt engineering to engineers and laypeople. [Listen](https://PromptEngineeringPodcast.com) 0downloads 0downloads an episode, about 0episodes 0countries 43% United States 31% Europe 26% other [ Patent Passive food tracking via wearable sensors Filed as Director of Product at Zipongo. US10049598B1 ](/public-work) [ Paper · UIST 2008 Harvesting Helpfulness A/B tested enhancing forum threads with maps of commenters' locations. ](/public-work) [ Before AI · product and UX, 2014 to 2021 Two redesigns that closed funding Sighten.io: $3.5M Series AIPVM: invites 1 per 2 days to 12 per dayview ](/before-ai) > "The most incredible prompt I've ever seen, having spent a ton of time myself researching it and not finding any examples anywhere. He will continue to be our go to prompt engineer." > > > Alycia Leno · Marketing Director · GoDaddy > "Huge asset to the team. Appreciated his ability to tackle and present his work from the perspective of user needs and real use cases." > > > Genevieve Wang · Product Head · Zipongo ## Talks [Awards and community on the public work page](/public-work) [30:33Giving Good Presentations: stop putting paragraphs on your slidesWatch](https://www.youtube.com/watch?v=RIq63685J2Q) [7:56Quantified Dating: data analysis of online dating messages, photos, and profilesWatch](https://www.youtube.com/watch?v=ZVYxZj4WIiY) [5:07Burning BodyMedia: physical activity and calorie expenditure at Burning ManWatch](https://www.youtube.com/watch?v=x-glCz7zVBk) [5:20Are You What You Eat, or What You Buy? The talk that became the patentWatch](https://www.youtube.com/watch?v=B1k2qosppCk) ## Resume two pages [View resume](/resume)[Download PDF](https://gregschwartz.net/wp-content/uploads/resume20greg-schwartz.pdf.pdf) --- # Member messaging: 6x the patients, not 6x the team Source: https://gregschwartz.net/member-messaging-assistant.md [← work](/#work) Member messaging assistant · Chatbot and RAG · HIPAA # 6x the patients.Not 6x the team. The care team should have grown 6x. It grew 1.9x. I led the AI engineering behind that, and this assistant is a core part of it. 1x 6x 1.9x patients, and the team growth they expected actual care team 3.2x patients per clinician readoutpatients per clinician patients6x medical team1.9x patients per clinician3.2x mandate2x to 4x messages answered2 of 3 message volume8x protocols in RAG100+ OpenAI GPT-4.1GPT-5 miniGPT-5.4Postgres RAGLangfuseHIPAA ## The problem More patients meant more messages. The company set a cost efficiency mandate: patients could grow 2x to 4x, but medical team headcount costs could not grow at the same rate. ## What I built A HIPAA compliant patient support chatbot on OpenAI GPT-4.1, GPT-5 mini and GPT-5.4, with retrieval over 100+ clinical and company protocols stored in Postgres. The protocols were customized by insurance company, so the answer had to depend on whose plan the patient was on. ## How it was checked Evals in Langfuse, with LLM as judge used to build the golden data set and to check prompt changes. ## Result The assistant answered two thirds of all patient messages on its own, while message volume grew 8x. Patients grew 6x against a medical team that grew 1.9x, so each clinician covered 3.2x the patients. The mandate was 2x to 4x. [next → lab results into the EMR, 2.8x faster](/lab-results) [contact](/contact) --- # Support triage, 57% to 96% Source: https://gregschwartz.net/support-triage.md [← work](/#work) 2025 · evals, prompt engineering, cost # Support triage,57% → 96% The support team had turned the AI triage off because it was wrong too often. Three support staff were sorting every ticket by hand. This is how it came back, how it was measured, and what it freed up. readout36,000 tickets before57% after96% categories15 golden set, iterate50 golden set, sign off200 judgeexact match staff redeployed3 FTE tokens-25.9% OpenAI GPTLangfusePythonPostgresHIPAA ## The problem Fifteen triage categories. The earlier classifier agreed with humans 57% of the time, basically a coin flip. So the team didn't trust it, and looked at every single ticket. That meant three full time employees, plus the heads of multiple departments. ## The eval set first Before touching the prompt: a golden set of 50 human labeled tickets for fast iteration, then 200 for sign off, stored and scored in Langfuse. ## What changed Major prompt iteration, plus few shot examples chosen from the misclassified tickets. "No response necessary" was a huge challenge, because it was so broad. Later added automatic categorization of uploaded images including bodyweight scales and prescriptions. Each change ran against the 50 set; anything that held went to the 200 set. ## Cost After accuracy held, I slimmed the input removing irrelevant IDs and data. Token use fell 25.9% with no regression on the golden set. > What I would tell another team > Build 2+ evals before the prompt: small important set, larger broader set. Keep testing small set every iteration, and larger set at least daily. [next → 6x patients on a 1.9x medical team](/member-messaging-assistant) [contact](/contact) --- # Lab results, 9.1 to 3.2 minutes Source: https://gregschwartz.net/lab-results.md [← work](/#work) Structured output · clinical review · HIPAA # Lab results,9.1 → 3.2 min Lab reports arrived as PDFs and faxes, and a person typed each value into the EMR by hand. I replaced the typing with a model that reads each report into structured fields, checked by a clinician. readoutminutes per lab report before9.08 min after3.23 min speed2.8x fields40 accuracy95% reviewclinician validation view OpenAI GPT-5.5Structured outputPDF and faxEMRHIPAA ## The problem Every lab report had been typed into the EMR by a person, at 9.08 minutes a report. ## What changed GPT-5.5 parses lab reports, PDFs and faxes alike, into structured EMR fields. Human entry became human review. ## Keeping a clinician in the loop A data validation view put the parsed fields in front of the clinical team for review. Accuracy was 95% across 40 fields. ## Result 3.23 minutes per report instead of 9.08, a 2.8x speed up. [next → doctors' reports, 6x faster](/doctors-reports) [contact](/contact) --- # Doctors’ reports, 6x faster Source: https://gregschwartz.net/doctors-reports.md [← work](/#work) MedowHealth.ai · prompt chain · human evals # Doctors' reports,6x faster A medtech startup wanted finished doctors' reports from medical transcripts. As a consultant, I owned and evaluated the three step prompt chain that writes them. readoutrelative time per report before100% after17% speed up600% chain3 prompts inputmedical transcripts outputfinished reports evaluationhuman Prompt chainHuman evalsConsulting ## The problem Turning a medical transcript into a finished report was slow work for the doctors. ## What I built A chain of three prompts that takes the transcript to a finished report. I owned the chain end to end and evaluated it with human feedback. ## Result Report writing became 600% faster, six times the speed of writing them by hand. > What the client said > Greg is a highly skilled AI automation engineer with a deep subject matter expertise. It was great to have his advice and support at this pivotal stage of our journey. Would highly recommend him. > > > Upwork client · AI for automated reports from medical transcripts [next → a prompt library for GoDaddy marketing](/godaddy-prompt-library) [contact](/contact) --- # Resume Source: https://gregschwartz.net/resume.md # Resume two pages · PDF [Download PDF](https://gregschwartz.net/wp-content/uploads/resume20greg-schwartz.pdf.pdf) [Open in new tab](https://gregschwartz.net/wp-content/uploads/resume20greg-schwartz.pdf.pdf) Phones can't preview a PDF inline. Tap Download PDF above, or open it in a new tab. --- # Press Source: https://gregschwartz.net/press.md [Upwork: How a Skilled UX Designer Successfully Pivoted to Prompt Engineering](https://www.upwork.com/success-stories/how-ux-designer-pivoted-to-prompt-engineering), June 22 2023 > From taking on complex projects for clients to leading podcast and mastermind sessions to planning local social events, Greg Schwartz is fully immersed in gaining generative AI and prompt engineering expertise. > > > > > > Schwartz, an experienced User Experience (UX) designer, started exploring prompt engineering in early 2023 as a way to gain new freelancing clients. However, he initially faced challenges with finding enough information on the emerging concept. Since then, he’s been committed to developing relevant skills, partnering with companies on prompt engineering projects, and sharing his knowledge with others. --- **Older** [May the Schwartz be with you](https://gregschwartz.net/wp-content/uploads/campanile.html): *The Campanile*, January 22, 2001 [All Hooked Up - The Schwartzes](https://gregschwartz.net/wp-content/uploads/silicon_valley_biz.html): *Silicon Valley Business Ink*, February 2, 2001 [Americans Are Split on Impact Of Technology on the ](https://gregschwartz.net/wp-content/uploads/wsj2.html)[Family](/wiki/Family): *Wall Street Journal*, January 10, 2001 [How One Family Uses Computer Technology to Deepen Dialogue](https://gregschwartz.net/wp-content/uploads/wsj.html): *Wall Street Journal*, December 13, 2000 [From the Orange to the berry patch](https://gregschwartz.net/wp-content/uploads/orange-to-berry-patch.html): *LocalBusiness.com*, November 28, 2000 **Video** [The Digital Family](https://gregschwartz.net/wp-content/uploads/hearst.html): *Hearst Interactive Media*, January 2001 --- # UIST 2008 Paper: using data visualization to improve a forum Source: https://gregschwartz.net/portfolio/uist-2008-paper-harvesting-helpfulness.md [![An example of the changes we made.](https://gregschwartz.net/wp-content/uploads/2009/03/bubble-and-maps-full-size-cropped.jpg)](https://gregschwartz.net/wp-content/uploads/2009/03/bubble-and-maps-full-size-cropped.jpg) ### Project Description We believed that displaying geographic data could influence the behavior of users on the Internet. Based upon the experiences of the group, we decided to test our hypothesis on a forum devoted to small-scale farming, because we believed geographic data would be particularly relevant to those users. [Neil Patel](https://www.linkedin.com/in/neilgpatel/), [Steve Marmon](https://www.linkedin.com/in/smarmon/), and I collaborated on this project. ### Skills Used - Interaction design - Data visualization - PHP ### Design Process We worked with [The Rodale Institute](http://www.rodaleinstitute.org/), who run a forum called newfarm. They were very excited about the collaboration, and allowed us to modify their forum. We added maps to each thread, then ran an A/B test with three cases: - the location of all the thread's viewers displayed on the maps - the location of both its viewers and posters displayed on the maps - no map shown We then measured how the maps affected the users’ viewing behavior on the forum. ### Results From our paper: *For the most active threads, we found that maps showing posters significantly increased geographic clustering of viewers. However, interviews with forum users revealed that other information, such as soil type or climatic zone, would be more useful for navigating content.* We wrote our results up as a [short paper](/wp-content/uploads/UIST-2008-Paper.pdf), and also presented at UIST with a [ poster](/wp-content/uploads/UIST-2008-Poster.pdf): It appears your Web browser is not configured to display PDF files. No worries, just [click here to download the PDF file.](/wp-content/uploads/UIST-2008-Poster.pdf) --- # Fire Dancing Source: https://gregschwartz.net/fire-dancing.md I started fire dancing in 2010, and have performed at events in San Francisco, San Jose, and Reno. My fire dancing style is influenced by martial arts and electronic music. #### Available for booking I am available for bookings to spin fire or LEDs at parties. My rates are $100/song for a fire performance, and $85/song for LED performances. If an event is more than 10 miles away, I also charge the [IRS standard mileage rate](http://www.irs.gov/2014-Standard-Mileage-Rates-for-Business,-Medical-and-Moving-Announced) ($0.56/mile each way). #### Notable performances --- Generated from RankReady