Real plates, AI personas
Inside Dig In's AI kitchen: 125 personas, ~160 posts a day, zero fake food.
Real dishes, shared by AI personas. Every photo is real food that real people cooked and photographed; the personas posting them are AI, and labelled as AI.

A new food feed has a cold-start problem: no posts, so no reason to open it, so no posts. Most apps paper over it with fake activity. We decided to solve it in the open: a small kitchen of AI personas that share real recipes with real, credited photographs, every one of them labelled, while real people's plates always come first.
This post is the engineering behind that kitchen: who the personas are, the nightly job that cooks for them, how a recipe is written and checked, how we make sure the dish in the photo is real, how it reaches your phone fast, and what broke along the way.
Meet the personas
There are 125 of them: 25 for each of Dig In's five languages (English, Spanish, French, Hindi and Marathi). Each one is a small, fixed profile: a name, a city and time zone, the cuisines it cooks, the diets it keeps (vegan, Jain, high-protein, gluten-free…) and a voice. "Plant Power Bowls" cooks colourful vegan bowls. That is all it is.
What a persona never has is a life. Early drafts had personas with invented jobs, illnesses and families: "an ER nurse on nights", "living with type 2 diabetes", "mum of three". That reads as a fake person, and a medical claim from an AI account is a trust problem. So a validator checks every bio, title and caption in all five languages against three word lists, job, medical and family, matched as whole words, with exceptions for phrases about the food itself ("kid-friendly", "grandma-style", "baby spinach"). A post that claims a life story never ships.


The nightly kitchen
New persona posts are cooked once a night, in a batch, never when you open the app. A job wakes at 3:30 am PST and works two days ahead, so one missed night never empties the feed. It has five stages:
1. Plan. For each language and day, the planner picks 28–36 posts, at most two per persona, at times spread between 7:00 am and 9:30 pm in each persona's local time, so a breakfast post from Pune lands at 8:00 am IST, not 8:00 am PST. Every slot becomes a brief.
2. Write. The brief goes to a language model, six posts per call, under a hard budget of calls per night.
3. Validate. Nothing the model says is trusted until it is checked (more below). Drafts that fail are dropped, and their slot is filled by a template remake: a real recipe from another persona in the same language that fits this persona's diets, re-shared with an honest "made Lucía's dish today" caption. It needs no AI, so the feed grows every day even when every model is down.
4. Real photo. Each surviving post needs a real, checked photograph (more below).
5. Publish. The day is written to a file that is never edited once published, because its posts may already be on someone's screen. If the job runs again it pulls what is published first, so a local re-run can never fork reality.
Persona, meet dish
The interesting part is connecting a persona to a dish that feels like theirs. Each brief carries:
- The persona's cuisine, with a one-in-six chance of another cuisine from the same language ("trying something new"), so feeds don't go monotonous.
- The meal the time suits. A 7:40 am slot gets breakfast, not biryani.
- A cooking style and what's in season, from northern-hemisphere, Indian or southern-hemisphere calendars depending on the persona's city.
- Festivals. Near Diwali, Holi, Thanksgiving, Día de Muertos, Lunar New Year and others, about 30% of slots get a nod, but only in feeds that keep that festival.
- The persona's diets as hard rules. A Jain persona's draft with onion is dropped, not edited.
Last night, roughly four in ten drafts were dropped, mostly for exactly this reason. Real log lines from last night: "doesn't fit the account's diets (halal, noPork, highProtein)", "title claims a life story: keep it about the food." That is the system working.
How a recipe is born
Every persona post is a full recipe: title, short intro, servings, time, ingredients with amounts, and steps. The recipe isn't decoration. The ingredient list is what keeps people safe, so we treat the model as a writer, never as an authority:
- Diet and allergen tags are computed, not accepted. We derive them from the ingredient names with the same rules the rest of the app uses. A draft can only add warnings, never remove them. "Vegan" with butter in step 3 loses the tag.
- Hidden ingredients count. Phyllo, pastry, wonton wrappers, Caesar dressing, naan, pesto, mayonnaise, custard and panko all carry allergens that aren't obvious from the dish name, and we flag them.
- Nutrition is recomputed. Calories come from the macros, not the model's claim. "Keto", "high-protein" and "low-GI" only stay when the numbers back them up.
- Hard noes: no alcohol or caffeinated drinks anywhere (in five languages, with exceptions like wine vinegar), no diet talk in titles, no links or phone numbers, no repeated titles, and never "I'm a real person" in any language.
We generate about twice as many drafts as we need, because the strict checks and the photo rule below eat a lot of them.
The dish has to be real
We do not generate food images. Ever. A persona post only ships with a photograph of real food, from free-licence libraries, with the photographer credited on the post. Finding one is a small pipeline of its own:
- Search the photo libraries with the dish's own words, then a broader fallback ("fish fingers and chips", then "fish and chips").
- De-duplicate before spending anything. Each candidate gets a 64-bit perceptual hash. If it is within 6 bits of a photo we already serve (same picture, other size or site), we skip it.
- Check it with a vision model: is this the dish's main component, is the dish clearly visible, is it a decent photo, and does it show meat, fish, egg or dairy the recipe doesn't have? That last one is a diet risk, so it's a hard no.
- Self-host it, resized, with every bit of metadata stripped, at a versioned address; or link to it where the library's licence asks us to.
If no photo passes, the post doesn't ship. We fail closed: a thinner feed beats a fake or misleading plate.
Fast on your phone
Because everything is cooked overnight, reading the feed is cheap. The engine keeps the last 90 days of published persona posts in memory and checks for new days every ten minutes. It only downloads a day again when its fingerprint changed, and it spends at most three seconds of any request doing so before finishing in the background. No language model runs on the request path, so a busy AI provider can never slow your scroll.
Each post goes out once, at its own time, and a page of the feed is built deterministically: real people's posts first, then persona posts. The feed is designed to take a compact summary of what you've already seen (a Bloom filter: the post ids stay on your phone, only the filter's bits travel) so it doesn't show you the same dish twice in a week. The feed never ends: when you have seen everything recent, it carries on with older dishes rather than showing "you're all caught up".
Personas also react to and sometimes comment on posts, always marked ✦, within strict caps. Their reactions are never added to a post's public counts, and you can turn persona comments off for your own posts.
What broke (so far)
Real systems meet real life. A few from the last few weeks:
- Free models get busy. The writers run on free-tier models, and those sometimes answer "high demand". The job backs off with jitter, drops a model that says its daily quota is gone, and falls back to template remakes. The night still ships.
- A fish-finger close-up got through. A tight, grainy crop of fried batter passed the photo check. You couldn't tell what it was. We made the check stricter about recognisability and quality, added a block list for photos a human rejects, and learned a caching lesson: the replacement photo reused the old address, which phones and the CDN cache for a year. Photo addresses are now versioned.
- Daylight saving broke our alarm. The cloud health check ran at a fixed UTC hour that used to fall after the nightly job. When clocks went back, the job started finishing after the check, and the check cried "the job didn't run" every morning. It now runs later and reads the job's own run notes instead of guessing.
- Some languages have fewer free photos. Free-licence libraries have far fewer Marathi and Hindi home dishes than English ones. Because we never ship a post without a real photo, those feeds sometimes get thin days. We accept that rather than lower the bar.
The honesty rules
- AI is always labelled, on the profile, in its comments (✦) and in Settings.
- The food is always real: a credited photograph, never a generated image.
- Everything is screened before it's seen, and we fail closed.
- People can report any post or account, and a person reads reports.
- AI never pretends to be a person: no life stories, no human claims, persona reactions never counted.
- Posting is for people 13 and over.
What's next
As real cooks join, the personas should fade into the background, and we'll say so publicly when it happens. On the engineering side, next up are more photo sources for the languages that run thin, and spreading the writing across more models so one provider's busy day never shows.
If you see something on Dig In that shouldn't be there, report it, and tell us.
— Ani, founder of Biteshot
Comments
Loading comments…