Perle

We are proud to share that our Founder & CEO, Ahmed Rashad , has joined the MENA to the Valley 2026 mentor network. Leveraging lessons from Scale AI, Amazon, and McKinsey, Ahmed will lead a session on Corporate Growth & Business Development to support the region’s next wave of tech founders. Full details in the post below. 👇 #Perle #MENAtoTheValley #StartupGrowth #BusinessDevelopment #Entrepreneurship

Most audio evaluations tell you if a model can transcribe words or classify a noise. Few tell you whether it can tell apart the sounds that matter — the difference between a wireless doorbell and a smoke alarm, or a police siren and an ambulance. We just released a benchmark on sound source identification: 11 methods — 4 Gemini models, the open-weight Kimi-Audio-7B, four traditional audio taggers, zero-shot CLAP, and an LLM-judged open-vocabulary model — tested on 2,242 real-world clips across 23 fine-grained sound classes. The core finding: Every method we tested is good at the broad category and bad at the specific sub-type. Models correctly hear "siren" or "running water" — then lose 25 to 50 F1 points trying to say which siren or how far away the water is. Traditional audio taggers even beat the best LLM at coarse classification without ever seeing the candidate list — but fall well behind once the task gets fine-grained. We also found none of the models show meaningful confidence calibration: wrong answers are phrased just as confidently as right ones, 92–100% of the time. You cannot use a model's tone to know when to trust it. If your product needs to distinguish specific sound sources — not just detect "a sound happened" — this gap is the thing to design around, not an edge case. Read the analysis: https://lnkd.in/eUK36vfa Test your model: https://lnkd.in/eKpx-uhY Full platform: https://perle.ai/benchmark #AudioAI #MultimodalAI #LLMEval #AIEvaluation #AcousticAI #ModelBenchmarking

Most audio evaluations tell you if a model can transcribe words or classify a noise. Few tell you whether it can tell apart the sounds that matter — the difference between a wireless doorbell and a smoke alarm, or a police siren and an ambulance. We just released a benchmark on sound source identification: 11 methods — 4 Gemini models, the open-weight Kimi-Audio-7B, four traditional audio taggers, zero-shot CLAP, and an LLM-judged open-vocabulary model — tested on 2,242 real-world clips across 23 fine-grained sound classes. The core finding: Every method we tested is good at the broad category and bad at the specific sub-type. Models correctly hear "siren" or "running water" — then lose 25 to 50 F1 points trying to say which siren or how far away the water is. Traditional audio taggers even beat the best LLM at coarse classification without ever seeing the candidate list — but fall well behind once the task gets fine-grained. We also found none of the models show meaningful confidence calibration: wrong answers are phrased just as confidently as right ones, 92–100% of the time. You cannot use a model's tone to know when to trust it. If your product needs to distinguish specific sound sources — not just detect "a sound happened" — this gap is the thing to design around, not an edge case. Read the analysis: https://lnkd.in/eUK36vfa Test your model: https://lnkd.in/eKpx-uhY Full platform: https://perle.ai/benchmark #AudioAI #MultimodalAI #LLMEval #AIEvaluation #AcousticAI #ModelBenchmarking

We build evaluation for the domains where standard metrics quietly lie to you: dialect, culture, law, medicine, agent behavior. Places where a model can sound fluent and still be wrong, and where that gap is exactly what breaks in production. Our newest one is public. It grades Arabic cultural and sociolinguistic knowledge. The finding: Implicit cultural reasoning, not vocabulary, is where models fail. And so do the automated graders. If you're using an LLM to judge quality, it's most confident in exactly the spots it's most wrong. How we build it: Prompts and rubrics written by subject-matter experts. Penalty-weighted scoring, so it's clear what has to be present and what counts as an error. Human ground truth first, automated judges second. Reporting that names who's accurate, who's lenient, and where each model breaks. We covered 103 expert-authored prompt and rubric pairs across Egyptian and Iraqi Arabic, grading frontier models both as answerers and as judges against native-speaker ground truth. If you're shipping into a high-stakes or multilingual market and need evaluation that holds up to expert scrutiny, we'll run it for your model, your language, your domain. Explore the results: https://lnkd.in/gP_PtKjk Paper: https://lnkd.in/gqmHfVCW Benchmark your model: https://perle.ai/benchmark #LLMEval #AIEvaluation #ArabicNLP #ModelBenchmarking

We built the first real benchmark for code-switched speech recognition, and it exposes something the demos hide: most commercial ASR breaks the moment a speaker switches languages mid-sentence. That isn't a rare case. It's how millions of people actually talk. What we did: 1,200 hand-verified code-switched utterances across Egyptian Arabic, Saudi Arabic, Persian, and German, each paired with English. Seven commercial systems, scored on WER and BERTScore. The takeaway: German and English is basically solved. Arabic and Persian are where systems fall apart, and where the ranking actually tells you which ones to trust in a multilingual market. If that's your market, this is the eval to look at instead of another clean single-language table. See the results: https://lnkd.in/gmPwTECJ Paper: https://lnkd.in/gcC-7PdJ Dataset: https://lnkd.in/gVqcNBmk Want your model on the board? We're taking requests. #ASR #LLMEval #ArabicNLP #CodeSwitching

We build evaluation for the domains where standard metrics quietly lie to you: dialect, culture, law, medicine, agent behavior. Places where a model can sound fluent and still be wrong, and where that gap is exactly what breaks in production. Our newest one is public. It grades Arabic cultural and sociolinguistic knowledge. The finding: Implicit cultural reasoning, not vocabulary, is where models fail. And so do the automated graders. If you're using an LLM to judge quality, it's most confident in exactly the spots it's most wrong. How we build it: Prompts and rubrics written by subject-matter experts. Penalty-weighted scoring, so it's clear what has to be present and what counts as an error. Human ground truth first, automated judges second. Reporting that names who's accurate, who's lenient, and where each model breaks. We covered 103 expert-authored prompt and rubric pairs across Egyptian and Iraqi Arabic, grading frontier models both as answerers and as judges against native-speaker ground truth. If you're shipping into a high-stakes or multilingual market and need evaluation that holds up to expert scrutiny, we'll run it for your model, your language, your domain. Explore the results: https://lnkd.in/gP_PtKjk Paper: https://lnkd.in/gqmHfVCW Benchmark your model: https://perle.ai/benchmark #LLMEval #AIEvaluation #ArabicNLP #ModelBenchmarking

Next up - we’re honored to welcome Ahmed Rashad, founder of Perle, to the inaugural Kernel Camp showcase. RSVP to hear from him! https://luma.com/igunc6l4 Before launching Perle in 2024 to build the foundation of AI, Ahmed led teams, product, and growth at Amazon and Scale AI. Ahmed will join Mona Ahmad and Ahmad Sadeddin for a panel discussion on MENA tech talent inside Silicon Valley startups, and why the region continues to produce globally competitive founders and engineers. Ahmed has been deeply involved in building across both the MENA and SF ecosystems, and we’re excited to have him share his perspective with the community. cc Hani Azzam Zaid Farekh

We're #hiring a new Expert Audio Transcriber, Tagalog (Filipino) in United States. Apply today or share this post with your network.

One of our founders helped Scale AI become the $29B gorilla in the room that led to Meta's $15B investment for 49%. He left and raised $17.5M to build his own human data company. His name is Ahmed Rashad. His startup is Perle. Fresh out of college, he started his career as an offshore oil driller. He went to work at Oracle, McKinsey & Company and Amazon. He joined the year Scale AI raised a $100M series C. He left the year after Scale raised $325M series E. In January 2024, he joined AI Circle as one of our founding members. In February 2024, he started Perle and raised a $8.5 pre-seed round. In August 2025, he raised a $9M seed round. $17.5M total funding. Tomorrow, we're going to host an AMA with Ahmed Rashad at Data Gradient. I'm grateful for Shauna Nehra at Cohere to fill in for me to facilitate the session. If there's one thing I've learned about Ahmed over the years, it's this: - He always delivers. - He has great war stories. - He gives great practical advice. - He has seen every iteration of human data. - He is all about giving his best no matter what he does. This session will be no exception. Sign up here: https://lnkd.in/eveaWVyZ