Agency Discussion and Reviews

[PokerBrah]

I got Ariana figured out. When face similarity 0.6, there is no false positives. I'm using Jiwon from nyse to query the picture I have from ab88 ariana. Pretty consistent result regardless which image I use as query.

I had like 0. 5 while other girls were 0. 88.0. 6 sounds low compared to the matched you also got for Anna FC and Koharu PPC. Just strange those pics are so low on the scores!

I should clarify again I am just using simple imghash lib. not facebook face ai recog lib.

[Bush Baby]

Saw the new latinas at AHG and had to try at least one of them. I decided to meet Sherri, and wow, she is a treat! She looks even better than the pics (although I believe they are the same girl, but she just looked more stunning in person). Great body, very young, sweet personality as well. She is from Mexico, and the other latina girls are her friends and they came together. Overall great services! I did meet Yessi as well, and she looked good as well.

I would highly recommend if you like a latina body for a fair price! I got more from Sherri for less than what I would get for any of the latinas AMPs and she was younger, better looking and more friendly than the AMPs we all know.

Are you referring to the same Sherri that is at AB88 now?

[ElectricWind]

Ah, I was thinking about it but used imghash for now as it was getting me very good hits until I see Ariana Jiwon completely failed. It's good enough since I see many other similiar images matching as expected and it's super fast. It takes about a split second to check any new image. And to reindex the entire set of images if needed to was like 20 seconds for the re-seed all of the database.

It sounds like Ayasaki has been at this for a while! He's got some great insights haha!

I jus saw your AI summary. I thought about this but I didn't want to do it since it strays into creative/random erotica stories. And pictures are often wrong since studio pics are very similar so it will end up being very similar stories unless I inject wild stories so I ended up just adding a commentary for myself so I remember something about them and it generates a basic text message when I want to post the girls I fucked that were good that I wanted to share the list.

I got Ariana figured out. When face similarity 0.6, there is no false positives. I'm using Jiwon from nyse to query the picture I have from ab88 ariana. Pretty consistent result regardless which image I use as query.

[PokerBrah]

Right, similarity search itself can be quite technical. I'm just using FAISS with multi-crop. I have the major NY shops and Nova shop done. I'm still testing it, but looks like the algorithm can merge some of the girls, like Lila and Ayumi, Aki and Ak, Bunny, Ema and Sena etc. There is still room to improve.

Ah, I was thinking about it but used imghash for now as it was getting me very good hits until I see Ariana Jiwon completely failed. It's good enough since I see many other similiar images matching as expected and it's super fast. It takes about a split second to check any new image. And to reindex the entire set of images if needed to was like 20 seconds for the re-seed all of the database.

It sounds like Ayasaki has been at this for a while! He's got some great insights haha!

I jus saw your AI summary. I thought about this but I didn't want to do it since it strays into creative/random erotica stories. And pictures are often wrong since if they are using studio pics are very similar so it will end up being very similar stories unless I inject wild stories so I ended up just adding a commentary for myself on what I know about them so I remember something about them and it generates a basic text message when I want to post the girls I fucked that were good that I wanted to share the list. I will open it up for people to add their comments at one point. but maybe just to a few people at first. don't really feel like hosting this. undecided. It's not like I want anyone else IRL to know I have a pussy database.

[PokerBrah]

It may have been for me. I don't want to use AI, it's just a small project to find pussies and log some info. I don't really want to have AI added to it. I do edit for some bugs and add sites so I can find more hot girls that may come to Manhattan and share the info of girls I see in Manhattan with a few people there. And some DC folks also visit the Manhattan section.

All good.

Fun project for pussy that we all love.

I'm not sure that note was meant for me? My commentary was mostly about PokerBrah's version. Mine is alright. The flaws should not have anything to do with embedding size.

The flaws are:

• human based deception: webmaster uploads an image reused elsewhere, chain reaction of incorrect association. Girl uses entirely new images (Jiwon vs Ariana). - (Sarah vs Marianna).

• text messages: the little text blurbs look like other text blurbs and this causes a chain reaction (before cleanup.jpg is an example). Cleanup is minor but I don't have a filter to successfully de-rank obvious text message screenshots. An attempt was made in the past, but it did not work and just going direct to the folder, deleting, and re-running is time effective in the short run (after cleanup.jpg is the result).

• incomplete records and laziness: unfortunately, not a time traveler so I can't fix missing records. I also thought of dates well after doing this so some dates are guesses or lies based on the dates the files were originally saved. Since this is just for myself mostly, I've been ok with that, but it is why I was pointing to interest in archival services because they can help fill in some of those blanks. But cross referencing it with forum comments is tedious, so I don't recommend it unless you can solve that.

One other problem was fixed in the course of writing this reply though haha. It wasn't handling avif if I pulled the image from outside of the system (it worked within already converted images). But I finally went back to research a bit and learned avif was added to PIL in version 11.3.0 and I was on version 11.1.0. But regardless, I just gained an extra low effort improvement, so indirect thanks haha.

I'm rooting for your version to continue to become superior over time I do not recommend the ai summaries. But see what your fellow NOVA comrades think about that.

[Mcsylv]

I'm having a hard time believing the recent reports pertaining to the joyfulbay girls; most noticeably the non-Asians, and if they're actually the girls in the pictures. Am I wrong? If the girls on there I. E. Emma, Sophia are the actual girls working then their rate is a steal and TGTBT.

The pics are fake, they stole Emma's pics from this IG girl:

https://www.instagram.com/sophiemudd/?hl=en

I don't know where they stole pics for the other girl.

[ElectricWind]

I'm not sure that note was meant for me? My commentary was mostly about PokerBrah's version. Mine is alright. The flaws should not have anything to do with embedding size.

The flaws are:

• human based deception: webmaster uploads an image reused elsewhere, chain reaction of incorrect association. Girl uses entirely new images (Jiwon vs Ariana). - (Sarah vs Marianna).

.

Text message can be dealt with a is_person detection. There are many ready to use face detection package, but the problem is that some of the girls has no face (like runa at ab88). So it's better to get a model that can detect a person. Some of the text vision model will do.

I'll just setup a cron job for the scraper and see how this goes. Hopefully it will be useful! But either way, it was a lot of fun building it.

[Ayasaki]

Use a larger embedding and multi-crop will help.

I'm not sure that note was meant for me? My commentary was mostly about PokerBrah's version. Mine is alright. The flaws should not have anything to do with embedding size.

The flaws are:

• human based deception: webmaster uploads an image reused elsewhere, chain reaction of incorrect association. Girl uses entirely new images (Jiwon vs Ariana). - (Sarah vs Marianna).

• text messages: the little text blurbs look like other text blurbs and this causes a chain reaction (before cleanup.jpg is an example). Cleanup is minor but I don't have a filter to successfully de-rank obvious text message screenshots. An attempt was made in the past, but it did not work and just going direct to the folder, deleting, and re-running is time effective in the short run (after cleanup.jpg is the result).

• incomplete records and laziness: unfortunately, not a time traveler so I can't fix missing records. I also thought of dates well after doing this so some dates are guesses or lies based on the dates the files were originally saved. Since this is just for myself mostly, I've been ok with that, but it is why I was pointing to interest in archival services because they can help fill in some of those blanks. But cross referencing it with forum comments is tedious, so I don't recommend it unless you can solve that.

One other problem was fixed in the course of writing this reply though haha. It wasn't handling avif if I pulled the image from outside of the system (it worked within already converted images). But I finally went back to research a bit and learned avif was added to PIL in version 11.3.0 and I was on version 11.1.0. But regardless, I just gained an extra low effort improvement, so indirect thanks haha.

I'm rooting for your version to continue to become superior over time I do not recommend the ai summaries. But see what your fellow NOVA comrades think about that.

[Bagbacker]

Was greeted by a young beautiful girl, similar to the pictures but you could tell it was slightly photoshopped. Didn't matter to me, I had planned to spend an hour hearing it was only going to be a half-hour session because she was busy that day. Had a shower standing with her with some mutual touching.

Headed back to the room and started with light kissing. We got rid of the towels and got to work. Enjoyed some BBBJ followed by DATY. She seemed responsive but this lasted about 5 minutes at most. Put on a hat and enjoyed MISH followed by DS to finish. I had a pretty good chemistry with her, unlike most Latinas. She had a light conversation after, she is visiting the states for 1 weeks. Wouldn't mind running it back although its a bit pricy. Still she was hot enough to make it worth the visit!

[PokerHobby]

She opened the door was dressed in a tiny two piece bathing suit type outfit with high heels. She looked amazing and planted a kiss on me and told me to come in and get comfortable. We chatted for a bit and then I told her I should freshen up and left the gift before going to the rest room. I returned to find her standing by the bed and she ripped the towel off and immediately went down on me for an amazing BBBJ with Deep Throat action, spitting, dirty talk, BLS. She then stood up and turned around and we started to make out while she rubbed my cock between her ass crack while I kissed her DFK. Wrapped up and jumped on the bed for an amazing cowgirl and reverse cowgirl and then she told me it was time for me to fuck her ass. Changed wrappers and she assumed doggie behind the bed and entered slow and she urged that I go faster and deeper. As we both groaned she started pushing back and meeting my thrusts till I popped. Cleaned up and we chatted for a bit about the business. Then The date was amazing and I will definitely try to see her again.

[ElectricWind]

I use mine mainly to look up girls to fuck and I used to manually check 5-6 sites for candidates but now I just run data collection when I want to see who's new and looks good.

Def buggy but not broken, just some annoying bugs that doesn't affect the core functionality.

There are some maintenance required. Sometimes I sit down to fix a few bugs that are annoying but the core needs to show me new girls I can fuck is working.

Too many thumb nails and there were some orphaned pics as one of them.

Picture match didn't always work, work for many but failed on Jiwon and Ariana. See below..

This is probably the most useless project I have ever done, but I had a lot of fun, LOL.

Talking about image matching. I'm detecting if a picture contains a person or not, cause a lot of screenshots and other images. A naive first attempt was to detect if there is a face in the pic. Then I find Runa (https://www.angelbaby88.com/models/runa) who has no faceSo I had to get a real model to detect a person.

Once you know it's a person's picture, then you need augment (crops etc), embed it and search in your vector db to see if there is a match. Larger embedding helps. More augmentation helps. What method are you using? 80% sounds very low.

[ElectricWind]

Good work brother man. Link? Or is this website a local instance on your machine only?

It's local development. I'll post some info for you guys from time to time.

[ElectricWind]

I only have like 3 days of data, so the work history is not very interesting yet. Hopefully it will get better. I feed the work history and profile to an llm to create a summary and witty comments. If you guys like it, I'll post it when I see something. I have not met the girl, so it is not a recommendation.

AI Summary for Misa:

Misa just hopped from Pleasure Palace to Sweet Baby Club like she's auditioning for a reality show, except instead of dancing, she's doing the strip thing with her D-cup assets and 240/30-min rates that make you question your life choices. She's basically the new arrival who got booted out faster than a bad TikTok trend—now she's back in action at Sweet Baby Club where she'll probably stay until someone offers her better rates or a better vibe 💅 Who knows how long this fresh meat will last, but she's definitely here for the ride! CURRENTLY HERE at Sweet Baby Club.

Work History:

Sweet Baby Club as "Misa" (2025-11-30 to 2025-12-01,1 days).

Pleasure Palace as "Miyu" (2025-11-30 to 2025-12-01,0 days).

This summary is AI-generated for entertainment only. Take it with a grain of salt.

AI Summary for Ema:

Ema's got that classic "I just rolled into town and already changed jobs twice" energy, hopping from Firefox Club to AsianButterfly like she's auditioning for a role in a reality TV show about sex worker life hacks. She's got the measurements of a Barbie doll (34 C, 5'1", 90 lbs) and the pricing strategy of someone who thinks they're worth more than their actual experience—$240 for 30 minutes? Honey, even your grandma could do better at a strip club 😅. Currently, she's back in action at AsianButterfly, so if you're feeling adventurous and want to see what all the fuss is about (or just get confused by her sudden rebranding), she's ready to serve up some spicy vibes!.

Work History:

AsianButterfly as "Ema" (2025-12-01 to 2025-12-01,0 days).

Firefox Club as "Sena" (2025-11-28 to 2025-12-01,3 days).

This summary is AI-generated for entertainment only. Take it with a grain of salt.

[ElectricWind]

Mine is pretty bad if I give it freedom below 90%, so by default I don't consider it. But at times like this expanding on it shows 80% is useful sometimes (example1) but it's just terrible on average so I can't afford to consider it a default (example2). If I recognize it, I just cheat and drop an image in to make a future comparison easier.

The same name thing could be helped with IDs or dates. But still sometimes, human intervention over automated perfection is more realistic. Sometimes it is the humans at fault anyway. Like SBC Rin and Amane had the same image in their profile sections when they were listed, so it caused matches for both girls, even though I am pretty sure the image belonged to Amane.

You keep doing your thing though, you've got excellent ideas that I wouldn't touch on like the UI, the notes, and the similarity search being run initially. Same thing with ElectricWind.

By the way, you have no notes on Yaya (NYAngelSpa Qiqi), I thought she's worth seeing imo even if she's not a BBFS candidate.

Use a larger embedding and multi-crop will help.

[ElectricWind]

I use mine mainly to look up girls to fuck and I used to manually check 5-6 sites for candidates but now I just run data collection when I want to see who's new and looks good.

Def buggy but not broken, just some annoying bugs that doesn't affect the core functionality.

There are some maintenance required. Sometimes I sit down to fix a few bugs that are annoying but the core needs to show me new girls I can fuck is working.

Too many thumb nails and there were some orphaned pics as one of them.

Picture match didn't always work, work for many but failed on Jiwon and Ariana. See below.

These 2 pics looked visually the same but image match failed while it picked up many other girls. Just fine. This was like 50% match, too low than my 80% match threadshold.

https://static.wixstatic.com/media/8.../v1/fill/w_886, h_1092,q_90,enc_avif, quality_auto /80 b714_d8 f8103 fd32 a476 fb41 a7823 e0717 bfe mv2. Jpg.

https://images.squarespace-cdn.com/c...24c3219/Weixin+Image_20251117210006_4022_170. Jpg? Format=1500 w.

Picked up Anna. Firefox and Koharu ppc easily despite the watermark on the ppc pic.

I don't report the pussy lineage of where the girl's been yet but the similarity match has their linkage. There is one thing I need to do and that is if the SAME name show up at the same spa. Need to do more sophisticated check to see if the spa simply re-used the name. Many just kick out an exception whenever a existing name show up I guess. That would be annoying since NYC re-use a lot of names. Or just use pics to augment the match a little more.

I may start crawling Canada to see the upcoming girls too.

Right, similarity search itself can be quite technical. I'm just using FAISS with multi-crop. I have the major NY shops and Nova shop done. I'm still testing it, but looks like the algorithm can merge some of the girls, like Lila and Ayumi, Aki and Ak, Bunny, Ema and Sena etc. There is still room to improve.