Amarbayar. A

Co-Founder at DuranAI | Voice AI | Realtime, On-Device TTS/ASR | Agentic RAG | Ex-Twitter SRE Lead • Amazon • GitLab • Hopper | Scaled Zaya, B2C, to 100k users

[Amarbayar. A @nomad_amaraa]

@JayHoovy Wait… You said: “the whole journey”, but then you say: “solid business, early traction”? Since when did “the whole journey” exclude the ideation, MVP, cold reach outs…? 😬

[Amarbayar. A @nomad_amaraa]

@ivanburazin @upfrontvc Even pre-seed VCs be like: “You don’t have PMF” but some founders raise seed stage with a slide: “We will get it to PMF”. What a world 😬

[Amarbayar. A @nomad_amaraa]

In 2015, my dad showed the first signs of Parkinson’s. What started as a small twitch in his toe slowly progressed over the next 11 years into something that affected his movement and daily life.

In the brain, there is a region called the substantia nigra. It produces dopamine, which directly impacts movement and muscle control. In Parkinson’s, this system breaks down, and symptoms begin to appear.

Today, an estimated 11 to 13 million people worldwide live with this condition.

But this story is not just about the patient. It is about the caregiver.

For the past 11 years, my mom has been there every single day. She never gave up. She searched for answers everywhere. Medical and non-medical. Eastern and Western. Books, research, online forums, conversations with others. She tried everything that could possibly help.

She is over 60, yet she taught herself how to search online, read, learn, and connect with people just to support my dad. There were many difficult moments, but she kept going. Watching her, I often wondered how much strength a person can truly have.

Because of her, my dad is still doing relatively well today. He can walk, go outside, and have conversations.

Personally, I spent months studying the brain and nervous system, reading, watching, and trying to understand Parkinson’s more deeply. I joined multiple Parkinson’s communities, researched @MichaelJFoxOrg @ParkinsonDotOrg @ParkinsonsUK @Merck @MerckFoundation @merckgroup @APDAParkinsons @APDA, supported some of them, and connected with doctors along the way.

One of them was the B1 Parkinson’s Association. I came across work by Antonio Costantini, who explored B1 therapy, and later Daphne Bryan, who continued that work and built a global community.

In that 14,000+ member community, one problem kept repeating. People and caregivers kept asking the same questions. About protocols, dosage, side effects, when to continue, when to stop. Doctors were answering everything manually, one by one.

So we decided to help.

Over the past 5 months, I worked closely with doctors to build an AI-powered Q&A system. It allows people to ask questions in their own language and receive accurate, source-backed answers prepared by medical professionals.

Today, we officially launched it on https://t.co/1cubrAc7fB

This system can now support tens of thousands of patients and caregivers who need reliable information.

And we are not stopping here.

We are building toward a voice-first future. No typing needed. Just speak in your own language.

Imagine your parent simply talking about how they feel. The system tracks it, identifies patterns, and helps caregivers understand what is happening. Doctor visits become structured, clear, and data-driven.

Technology is moving fast, but many of our parents and grandparents are being left behind.

We are building to change that at https://t.co/V33furm280.

[Amarbayar. A @nomad_amaraa]

Organizing DevSummit (the largest tech conference in Mongolia) for the 5th year: https://t.co/XEwO0h3Srq. It is a 2-day event 5/30 - 5/31 where we will bring 1000+ engineers from Mongolian tech community for tech talks, panel discussions and workshops.

We are already in talks with western tech companies to see if we can bring some speakers. But would love to bring speakers from the following companies. What is the best way to reach these companies?

@AnthropicAI @GoogleDeepMind @Alibaba_Qwen @deepseek_ai @MiniMax_AI @OpenAI @AWS_Gov @awscloud @Microsoft

[Amarbayar. A @nomad_amaraa]

@andrewchen Thanks! Please have a look at https://t.co/wHJoNKi5Sc (a voice AI for caregivers and patients with Parkinson’s to help with daily check-ins, symptom tracking and correlation analysis. Built for my dad, now expanding to millions more people. Applied on 4/10th)

[Amarbayar. A @nomad_amaraa]

Hey @OfficialLoganK - Love this update! 2 questions I was curious if you can help answer:

1. Will you support streaming or bring it to the Live?
2. If you run the exact same inference multiple times you get different sounding voices though the voice, direction, style, past have't changed at all. Is this expected?🤔

[Amarbayar. A @nomad_amaraa]

I was trying to integrate with @HeyGen's APIs and test video generations were getting stuck. The Support asked if I tried cancelling and resubmitting the stuck video. With no proper SLA and instructions on how long to monitor for stuck requests before issuing "cancel" and "resubmit" it feels very manual to do so and kind of defeats the purpose of API integration. But then I noticed Claude was down... I suspect HeyGen is using @claudeai behind the scenes and is now impacted. Claude being down is likely causing tons of support tickets and inbound calls to businesses that wrap their APIs. One strategy businesses should adopt is multi-LLM provider integration to avoid such a single point of failure scenario. :/

[Amarbayar. A @nomad_amaraa]

@pmitu So true.

[Amarbayar. A @nomad_amaraa]

"GStack thinks you are among the top people who could do this" showed up on 5/5 projects I have built with @garrytan 's gstack so far... 😁

[Amarbayar. A @nomad_amaraa]

Voice AI companion for Caregivers and Patients with Parkinson’s — daily check-ins, med reminders and doctor visit ready reports. 11M+ patients and 50M+ caregivers globally. I have been a caregiver for my father, 11 years. Iterating fast with my dad and family using it and providing feedback. In talks with a doctor to bring this voice capability to their patient community. Product: Noongil, web: https://t.co/V33furm280 Happy to meet for 5 mins to show a demo and go over the product/vision.

[Amarbayar. A @nomad_amaraa]

Open source project: event-os
GitHub: https://t.co/XSthJs1zfC

Every year organizing DevSummit (the biggest Mongolian tech conference) means chaos across spreadsheets, Viber/WhatsApp/Messenger threads, emails, and scattered Google Drive folders.

Program, speakers, panelists, workshops, venue, sponsors, catering, design, invites, booths, volunteers, marketing/content, ticketing, check-in… plus task management across the entire team.

We’ve managed to pull off 4 events like this. For the 5th year, I want to see if we can simplify everything.

Launching event-os (experimental) — an open-source system to run events end-to-end.

Core ideas:
- 🏛️ Start with CRUD → evolve into agentic workflows that process inputs dynamically
- 🦀 Run on OpenClaw + integrate with Telegram so organizers can manage everything via chat (status, updates, coordination)
- 📊 Dashboard + RBAC for visibility and control

Goal: from event creation → scheduling → team coordination → ticketing → check-in → automated speaker notifications — all in one system.

Building this 100% vibe-coded and iterating in the open. If you’re interested, jump in and open a PR.

[Amarbayar. A @nomad_amaraa]

@brettcalhounn Hey Brett - I DM'd on 2/26 and haven't heard back yet. Is it still open? :)

[Amarbayar. A @nomad_amaraa]

We have been training an Automated Speech Recognition (ASR) model in low resource languages such as Mongolian and optimizing them to run on offline / CPU / realtime. We are turning some of the models into products so that users can get immediate benefit out of them.

Today, we are launching https://t.co/iyxGoyh3ag, a Mongolian meeting note taker app. The underlying ASR model achieves the following:

- Up to 91% accuracy
- Code switching between MN and EN
- The CPU version achieves RTF of 0.05x

[Mongolian]
📝 Хурлын тэмдэглэл хөтөлнө гэдэг бас л амаргүй ажил байдаг даа тээ? Дутуу сонсчих вий, буруу ойлгочих вий гээд л. Олон улсад Англи хэл дээр бол бараг онлайн хурал болгон л тэмдэглэл хөтлөгчтэй болж, хурал дууслаа бол тэмдэглэл, тойм нь бэлэн болчихдог. Монгол хэл дээр тийм юм бараг байхгүй.

📣 Ингээд манай DuranAI баг өнгөрсөн хэдэн сарын туршид Монгол хэлний яриаг текст болгон хөрвүүлэх чадвартай "ДЭЛДЭН" хэмээх модель дээр ажиллаж дуусаад, үүн дээр суурилж анхны бүтээгдэхүүн болох 🐰 Дэлдэн - Хурлын Тэмдэглэл Хөтлөгчөө танилцуулж байна.

🗓️ Google Meet, Teams, Zoom зэрэг онлайн хурал хийдэг хүн бүхэнд хэрэгтэй программ. Хурал дээр хэн юу хэллээ, ямар ямар шийдвэр гаргалаа, дараагийн алхмууд / ажлын жагсаалт юу билээ гэж санаа зовохгүй хурлаа хийхэд л болно. Дэлдэн энэ бүхнийг нь сонсоод, хөрвүүлээд гаргаад өгнө гэсэн үг. Зарим нэг онцлог чадварууд:

- Монголчууд маань Англи үг/өгүүлбэр их хольж ярьдаг. Гадаадын хүмүүстэй ч их уулзалт хийдэг. Дэлдэн codeswitch буюу Монгол <-> Англи хэл аль алийг нь сонсох чадамжтай.
- Таны тэмдэглэлийг зөвхөн та л харах боломжтой. Та өөрийн сонгосон нууц үгээр бүх хурлын тэмдэглэлээ шифрлэх боломжтой. Тэгснээр манай системд хадгалагдахдаа ч хүртэл шифрлэгдсэн байна. Зөвхөн та систем дотроо нууц үгээ хийсэн тохиолдолд л танд харагдана.
- Мэйлээр хурлын тэмдэглэлээ хүлээн авах боломж. Мөн PDF ээр татан авах боломжтой.
- Microsoft Teams, Google Meet, Zoom дэмжинэ. Google Calendar, Microsoft Outlook аль аль календарьтай холбох боломжтой.
- Календарь дээр байгаа видео линктэй хурлууд шууд дэлдэн рүү автоматаар орж ирэх бөгөөд, яаралтай шуурхай хурал хийх хэрэгтэй бол хурлын линкээ оруулахад л дэлдэн гүйгээд орно.

🎁 За тэгээд СЮРПРИЗ нь гэвэл манай багийн зүгээс ТА БҮХЭНД 1 ЦАГИЙН ҮНЭГҮЙ ашиглах бэлгийг өгч байгаа юм. Хувь хүн, байгууллага хамт олноороо хэрэглээд туршаад үзээрэй. Засаж сайжруулах зүйлс байвал бидэнд сэтгэгдэл, шүүмжээ хуваалцаарай. Харин таалагдвал цааш нь түгээж өгөөд, төлбөрөө төлөөд хэрэглээрэй 😆

Та Хуралдаа л Анхаар. Бидэнд Тэмдэглэлээ Даатга.

[Amarbayar. A @nomad_amaraa]

@garrytan @garryslist Why so big on Ruby? Are you testing your idea of Ruby On Rails potentially saving tokens because of the rigid structure it already has?

[Amarbayar. A @nomad_amaraa]

@sama After 5 rounds on Enterprise plan --&gt;
"You've hit your usage limit. Try again at 6:56 PM" 🤯

[Amarbayar. A @nomad_amaraa]

We recently trained a CPU optimized ASR model on top of a CPU optimized TTS model we had already trained. They achieve RTF of 0.05-0.08x.

We have experimented with the idea of realtime dubbing on Youtube videos and it works actually pretty good! We then made it better by adding speaker diarization and gender classification. Adds a bit of latency but now I can watch any Youtube contents in my preferred languages. I tried watching @netflix Physical Asia’s participants’ interviews in English (original contents in Korean and Japanese).

How is the dubbing industry? Heard @HeyGen is providing free dubbing to boost their Avatar usage (but consumers know their price is just too high), and @YouTube started rolling out free dubbing feature targeted towards @YouTubeCreators but it is very limited (and not great quality).

[Amarbayar. A @nomad_amaraa]

@andrewchen Only students &amp; recent grads? 🤔

[Amarbayar. A @nomad_amaraa]

@YuanAI_Lab 1T parameter for a RAG optimized model? Why 🤯

[Amarbayar. A @nomad_amaraa]

We benchmarked 10 LLMs (open source and commercial) on our multilingul, agentic information retrieval use case and compared:

- Factual Accuracy
- Coverage
- Latency
- Tokens spent / $ spent
- Number of tool calls

We prepared medical dataset (English and Mongolian) and also an insurance service dataset (English and Mongolian). And tested EN->EN, MN->MN, EN->MN, MN->EN, and transliterated MN (think of “Monglish”) -> EN routes.

Top 2:
- @GeminiApp Gemini 3 Flash
- @Zai_org GLM-5
surprisingly GLM-5 did very well, though it was pricier. Just connected with the https://t.co/rMkncRdAyt team and they are going to give us free API credits so we can push GLM-5 further to see how it does on more complex use cases. Thank you! 🔥

I am trying to find someone who can endorse on @arxiv so I can publish this paper, too. I think it has really useful data and others won’t have to spend days doing similar benchmarking 10 LLMs for agentic information retrieval but use our research for their needs.

[Z.ai @Zai_org]

https://t.co/DRCwFHQQRn Startup Program is NOW OPEN.

What you can get:
·Free API credits
·Priority rate limits
·Exclusive Community
·Early API Access

Who we're looking for:
·AI-native startups
·Agent builders
·SaaS founders integrating LLM infra
·Global teams building for real-world scale

If you're building something that matters, don't wait!!

Apply now: https://t.co/jyS5zVZKQx
Questions? Details? Follow & DM @ZaiforStartups