James Bradford

Most ai chatbots' conversations are very hard to get right and they bloat as context grows. I think its because most builders treat conversations like a straight line. To ACTUALLY build high level agents, you need to master conversational branching logic. Here is the blueprint we use at @akapulu_ai to build these chatbots at a high level🧵:

What are your guys' thoughts on youtube long-form, technical tutorials?

Got a black eye in jiu jitsu yesterday and i was actually able to use it for my AI avatar tutorial today. Helped show the vision function while booking my appointment lol Full tutorial drops Friday!

Cant wait to show you guys what we've been building.

'The future of avatars is intelligence, not just animation' Infosys, April 2026. Spot on!!

Dropping a new tutorial this week , building a healthcare intake AI avatar with @akapulu_ai . Vision, real appointment booking , full transcript analysis, all in about 20 minutes or so

Changing environments temporarily has been the greatest productivty hack ive experienced haha. Visted my sister out in utah and ive been here for almost a week. Nothing to do except cook, workout, and grind🥀

Built this with Claude Code + Livekit agent Skills cool to see self-fixing errors and working voice agent in 15 min. The skill vs MCP distinction is a game changer for voice agent dev

New tutorial: i got claude code to build a voice agent 40x better using LiveKit's new Agent Skills. 2% success from scratch → 80% with the right setup according to this blog: livekit.com/blog/voice-age… Full build in 27 min plus why single-prompt agents break at scale. youtu.be/92_H7pTf22k

Beta delayed, will's fixing npm hell, i'm on vacation pretending i'm not checking discord

Best ai agents ive used in prod are the ones that dont have much freedom Like heavily constrained as opposed to being able to do whatever they want Freedom is a UX problem in disguise lowkey

@Alfonso1127041 Spot on bro

if your setup is just waiting for inbound traffic to answer basic FAQs, you aren't using AI. you're just using a slightly better search bar

Most founders still think an AI sales agent is just a glorified chatbot widget on their pricing page. they're completely missing the shift.

@erogol Heavy TTS week indeed. The gap between raw TTS quality and actually deploying it is closing fast haha

Machine Learns - 67 🤖 Heavy TTS week + papers on self-improvement. • Voxtral-4B • Aurora: speculative decoding that learns from live traffic • Mamba-3 • Hyperagents: agents with metacognition • λ-RLM • TurboQuant • Directional Routing erogol.substack.com/p/machine-lear…

@agenticgirl Wild benchmark numbers. Curious how it holds up in a live conversation loop. Synthesis + latency is where it gets real, i assume it would do really well!

Mistral just dropped a TTS model and it’s beating ElevenLabs. Voxtral TTS. 4B params. Clones any voice from 3 seconds. 68.4% win rate over ElevenLabs Flash v2.5 in blind tests. → Autoregressive model handles what is said → Flow-matching model handles how it sounds That’s why it doesn’t just sound clean it actually sounds like the person. → 9 languages (incl. Arabic, Hindi, Spanish, Portuguese) → Stronger speaker similarity across languages → Works with just 3s–30s voice prompts → Sub-second latency (real-time feel) → 32 concurrent streams on a single GPU → Handles long audio without breaking context → Better expressivity (tone, emotion, style) → Open weights (CC BY-NC) → Beats ElevenLabs in voice cloning + competitive on flagship voices Most TTS models give you a voice. This gives you a person. Feels like a shift. Voice = identity. And now it’s open.

Guys there are two 'ai avatar' markets that i think people combine but theyre really separate Type 1: avatar renders video from your script. Type 2: Avatar holds a live conversation in real time. Different tech and different buyers, bc theyre fundamentally different problems Important to know which one youre building

Weve documented this mini philosophy and the specific node structures at: 👉 docs.akapulu.com If you want to see me actually build these live feel free to check the YouTube channel @JamesAkapulu Follow for more deep dives into this type of stuff!

Id say majority of builders only design the "Happy Path" (the perfect conversation) But in reality you need to build for "Detours." Eg recursive fallbacks. If the AI doesn't understand the intent after 2 tries, the branch logic should automatically pivot to a human handoff or a simplified menu.

Most ai chatbots' conversations are very hard to get right and they bloat as context grows. I think its because most builders treat conversations like a straight line. To ACTUALLY build high level agents, you need to master conversational branching logic. Here is the blueprint we use at @akapulu_ai to build these chatbots at a high level🧵:

Complex branching logic is the difference between a toy and a tool. When you solve for intent-switching and state-retention, it feels a lot less robotic and more like the (cliche i know) "digital employee" .