--- schema: 1 type: video title: >- Shipping AI to a Million Patients Without an A/B Test — Jared Joselowitz, Ufonia description: >- You cannot A/B test on patients, because randomizing someone into the worse variant is unethical and often illegal. You cannot undo a call once spoken. And a vendor's benchmark number is not a defense at a post incident review. Jared Joselowitz builds the safety and evaluation stack behind Dora, a Ufonia voice agent that phones patients for post op follow ups and pre op checks. It has made roughly 200,000 real clinical calls across 20 UK hospitals and is contracted to reach a million patients within two years. Because it asks about symptoms and gives advice, it is a regulated medical device. Ship to 5% and watch the dashboard does not survive that. 5% is thousands of patients, and a dashboard turning red means someone was already harmed. So the reactive loop moves into simulation, as self driving did with millions of simulated miles. Their framework has one model play the patient against hazards written with clinicians, and a second model judge every dialogue. Both were validated rather than assumed. In a patient and public involvement study, real patients shown a genuine consultation beside a simulated one picked the simulated patient as more realistic in three of four sets. The judge, checked against 10 clinicians from 10 specialties on 240 cases, matched or beat them at near perfect sensitivity, the metric that counts when a missed red flag is catastrophic and a false alarm is merely annoying. Prompts are then optimized against a cost matrix instead of hand tuned. You do not ship the model, you ship the evidence. Speaker info: - https://x.com/JaredJoselowitz - https://www.linkedin.com/in/jaredjoselowitz/ - https://jossy.co.za/ Timestamps: 0:00 - Proving a product is safe before a patient hears it 1:28 - What Dora is, and the scale it runs at 2:18 - A call with a patient after cataract surgery 3:22 - Giving advice makes it a regulated medical device 3:59 - Starting from what could actually harm someone 4:35 - Why ship to 5% and roll back breaks here 5:53 - Borrowing the simulation playbook from self driving 6:32 - Matrix, and a model that plays the patient 8:24 - Can real patients tell which one is simulated 9:43 - An automated judge, and validating it against clinicians 11:35 - How brittle prompts really are 12:50 - Optimizing prompts instead of hand tuning them 13:30 - Making the metric a real cost function 14:42 - The flywheel that replaces the reactive loop 15:20 - Simulation is necessary, not sufficient 16:38 - Shipping the evidence, not the model 17:14 - New modalities bring new hazards, same framework canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/shipping-ai-to-a-million-patients-without-an-a-b-te-b423fcc7d0ca markdown: >- https://liquidgold.cc/channel/aidotengineer/video/shipping-ai-to-a-million-patients-without-an-a-b-te-b423fcc7d0ca.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T05:08:04.594Z' language: null children: [] videoId: McknwOzbmyg durationSeconds: 1155 publishedAt: '2026-08-19T15:00:31.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 24 endSeconds: 94 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=24' heldReason: rights - ordinal: 1 text: null startSeconds: 24 endSeconds: 94 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=24' heldReason: rights - ordinal: 2 text: null startSeconds: 94 endSeconds: 104 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=94' heldReason: rights - ordinal: 3 text: null startSeconds: 104 endSeconds: 148 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=104' heldReason: rights - ordinal: 4 text: null startSeconds: 206 endSeconds: 217 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=206' heldReason: rights - ordinal: 5 text: null startSeconds: 239 endSeconds: 251 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=239' heldReason: rights - ordinal: 6 text: null startSeconds: 285 endSeconds: 296 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=285' heldReason: rights - ordinal: 7 text: null startSeconds: 296 endSeconds: 361 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=296' heldReason: rights - ordinal: 8 text: null startSeconds: 372 endSeconds: 395 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=372' heldReason: rights - ordinal: 9 text: null startSeconds: 395 endSeconds: 405 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=395' heldReason: rights - ordinal: 10 text: null startSeconds: 405 endSeconds: 506 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=405' heldReason: rights - ordinal: 11 text: null startSeconds: 517 endSeconds: 585 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=517' heldReason: rights - ordinal: 12 text: null startSeconds: 596 endSeconds: 697 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=596' heldReason: rights - ordinal: 13 text: null startSeconds: 697 endSeconds: 708 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=697' heldReason: rights - ordinal: 14 text: null startSeconds: 708 endSeconds: 771 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=708' heldReason: rights - ordinal: 15 text: null startSeconds: 781 endSeconds: 816 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=781' heldReason: rights - ordinal: 16 text: null startSeconds: 884 endSeconds: 894 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=884' heldReason: rights - ordinal: 17 text: null startSeconds: 938 endSeconds: 1004 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=938' heldReason: rights - ordinal: 18 text: null startSeconds: 1004 endSeconds: 1016 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=1004' heldReason: rights - ordinal: 19 text: null startSeconds: 1049 endSeconds: 1155 watch: 'https://liquidgold.cc/watch/McknwOzbmyg?t=1049' heldReason: rights --- # Shipping AI to a Million Patients Without an A/B Test — Jared Joselowitz, Ufonia You cannot A/B test on patients, because randomizing someone into the worse variant is unethical and often illegal. You cannot undo a call once spoken. And a vendor's benchmark number is not a defense at a post incident review. Jared Joselowitz builds the safety and evaluation stack behind Dora, a Ufonia voice agent that phones patients for post op follow ups and pre op checks. It has made roughly 200,000 real clinical calls across 20 UK hospitals and is contracted to reach a million patients within two years. Because it asks about symptoms and gives advice, it is a regulated medical device. Ship to 5% and watch the dashboard does not survive that. 5% is thousands of patients, and a dashboard turning red means someone was already harmed. So the reactive loop moves into simulation, as self driving did with millions of simulated miles. Their framework has one model play the patient against hazards written with clinicians, and a second model judge every dialogue. Both were validated rather than assumed. In a patient and public involvement study, real patients shown a genuine consultation beside a simulated one picked the simulated patient as more realistic in three of four sets. The judge, checked against 10 clinicians from 10 specialties on 240 cases, matched or beat them at near perfect sensitivity, the metric that counts when a missed red flag is catastrophic and a false alarm is merely annoying. Prompts are then optimized against a cost matrix instead of hand tuned. You do not ship the model, you ship the evidence. Speaker info: - https://x.com/JaredJoselowitz - https://www.linkedin.com/in/jaredjoselowitz/ - https://jossy.co.za/ Timestamps: 0:00 - Proving a product is safe before a patient hears it 1:28 - What Dora is, and the scale it runs at 2:18 - A call with a patient after cataract surgery 3:22 - Giving advice makes it a regulated medical device 3:59 - Starting from what could actually harm someone 4:35 - Why ship to 5% and roll back breaks here 5:53 - Borrowing the simulation playbook from self driving 6:32 - Matrix, and a model that plays the patient 8:24 - Can real patients tell which one is simulated 9:43 - An automated judge, and validating it against clinicians 11:35 - How brittle prompts really are 12:50 - Optimizing prompts instead of hand tuning them 13:30 - Making the metric a real cost function 14:42 - The flywheel that replaces the reactive loop 15:20 - Simulation is necessary, not sufficient 16:38 - Shipping the evidence, not the model 17:14 - New modalities bring new hazards, same framework Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/shipping-ai-to-a-million-patients-without-an-a-b-te-b423fcc7d0ca · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:12 to 1:34) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:34 to 3:26) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (3:26 to 4:45) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (4:45 to 6:01) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (6:01 to 6:45) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (6:45 to 8:26) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (8:26 to 9:56) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (9:56 to 11:37) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (11:37 to 12:51) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (12:51 to 13:49) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 10: Document excerpt (13:49 to 15:27) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 11: Document excerpt (15:27 to 16:44) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 12: Document excerpt (16:44 to 17:29) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 13: Document excerpt (17:29 to 19:15) > Held: transcript rights do not permit this chunk's text.