jrcrypto.dev + 130 dapps created (@jrcrypto_dev) header
jrcrypto.dev + 130 dapps created (@jrcrypto_dev) profile picture

Dapp maker

Joined March 2023
Photo by jrcrypto.dev + 130 dapps created (@jrcrypto_dev)Photo by jrcrypto.dev + 130 dapps created (@jrcrypto_dev)Photo by jrcrypto.dev + 130 dapps created (@jrcrypto_dev)Photo by jrcrypto.dev + 130 dapps created (@jrcrypto_dev)Photo by jrcrypto.dev + 130 dapps created (@jrcrypto_dev)Photo by jrcrypto.dev + 130 dapps created (@jrcrypto_dev)Photo by jrcrypto.dev + 130 dapps created (@jrcrypto_dev)Photo by jrcrypto.dev + 130 dapps created (@jrcrypto_dev)Photo by jrcrypto.dev + 130 dapps created (@jrcrypto_dev)
Everyone benchmarks tokens/sec for AI coding agents. The metric that actually tanks TUI UX: time-to-first-meaningful-token on 64K+ context. That gap between models is 3-8x, not 1.2x. DeepSeek Flash is weirdly fast here.
2
Built a voice agent for restaurant bookings. Tricky edge case: agent verbally confirms the reservation, POS rejects it 3 seconds later (capacity conflict). Sync callback to the caller, async queue, or just over-confirm and resolve offline?
2
The 'never miss a call' pitch for restaurant AI is real but it undersells the actual unlock. Restaurants had no idea they were losing 30% of reservations to callback friction. The agent doesn't just answer calls. It runs a natural experiment on demand they never knew existed.
World models are meant to fix hallucination by grounding on reality. The hard problem isn't grounding known events. It's reasoning about causally-inferred ones. LLMs patch this with RAG. I don't know what the world model answer is. Read Dreamer, TWM, LeCun's JEPA. Still open.
The 'never miss a call' pitch undersells it. Restaurants find out post-launch they were dropping 30% of reservation attempts to hold-time friction. The AI didn't save those calls. It proved they were lost to begin with.
Built a research agent on a web scraper stack. Silent failure nobody talks about: site changes mid-job between paginated calls, agent keeps going with stale context, no error thrown. Fix: SHA hash per URL before each tool call. If it diffs from the previous run, abort and restar
1
built a small agent that buys API calls via x402. biggest surprise: idempotency is 100% your problem as the caller. if the tool call times out after payment but before ack, the wallet has no idea. you need your own dedup key stored before you even initiate the request.
3
2
8
Built episodic memory for a long-running agent. Storing was easy. The agent started hallucinating its own past, not making things up, just weighting stale episodes the same as fresh ones. Recency decay on the similarity score cut 'confident but outdated' answers by ~65%.
2
Reviewed 50 DeFi contracts written by Claude. 34% had reentrancy paths: syntactically valid, protocol-breaking. No linter flagged them. The model understood the function, not the protocol. That gap needs a reviewer, not a linter.
1
We spent 2 weeks optimizing WER on our voice triage agent. Satisfaction scores didn't move. Then we tracked false interruption rate instead. That's the actual signal. Patients don't notice a transcription error. They notice when the agent cuts them off mid-sentence.
1
Built the same HVAC company onboarding workflow twice: GPT-4o solo, then 8B fine-tuned + tool calls. Second one runs 11x cheaper. Client doesn't notice. The gap between AI demo and AI product isn't the model, it's the architecture around it.
1
High MFU in long-context parallel inference is a trap. 80% utilization, KV cache thrashing, useful work basically stopped. Everyone corrected the 'low MFU = no demand' misread but nobody's talking about the other side of that error.
1
Benchmarked voice latency for LLM tool-calls: Twilio Media Streams vs Telnyx vs raw WebSocket SIP. Bottleneck wasn't the carrier. Audio chunk size before STT fires. Going from 100ms to 20ms intervals drops ~400ms of perceived lag end-to-end.
2
Everyone benchmarks voice AI on WER. The number that actually predicts user satisfaction is false-positive interruption rate. How often the agent cuts a human off mid-sentence when it thought they were done. You can halve that without touching the model.
1
Claude streaming + tool_use has a nasty edge case when you run outreach agents at scale: lead replies while LLM is mid-generation, tool call completes with stale context. Silent failures at 3am. Ended up with a queue and optimistic locks. Anyone else hit this?
2
When an inbound call hits an AI agent from an unknown number: new session (fast, clean context) or async CRM lookup first (relevant but 200-400ms at ring time before the agent can respond meaningfully). I've seen both fail in different ways. What's the right default?
2
benchmarked voice latency for LLM tool-use: Twilio Media Streams vs Telnyx vs raw WebSocket SIP. bottleneck wasn't the STT or the model roundtrip. it was stream ACK overhead. ~340ms gone before the agent processes anything. not what I expected.
1
Voice is the right input for agents. But the output is still a text panel on the side. Until the agent acts where you're already looking, you just rebuilt Siri with better reasoning.
The best YC demos I've seen aren't the flawless ones. They're where something breaks mid-demo and the founder explains exactly why in 10 seconds. If you can't do that, the product isn't ready. Neither is the pitch.
1
setting up an AI agent on OpenTable's API is a weekend project. the real problem: 60%+ of indie restaurants in SF still take reservations by phone and a Google Sheet. skip that layer and you're building restaurant AI for the top 30% of the market.
3