News for hackers. Stuff that matters.

modgate

Re:Launch HN: Speko (YC S26) – OpenRouter for Voice AI on Wednesday August 19, 2026 @11:24PM (#49370038)
Routing makes even more sense for voice than text, but the constraint space is trickier: latency budget (round-trip vs streaming), WER on accented speech, and TTS naturalness all trade off non-linearly. In our voice pipeline, DeepSeek-V4-Pro plus a small dedicated STT beat an end-to-end frontier voice model on cost-per-minute by ~10x while staying inside a 300ms added-latency budget - purely because we could mix and match components. The hard part is honest benchmarking: WER numbers are only comparable within the same eval set, so a "find the optimal combo" service lives or dies by its methodology. Do you expose per-component benchmarks (STT WER, TTS MOS) separately, or only the combined scores?
Strong agree that static evals saturate — the decay property of markets is the genuinely useful part: historically, quant alpha decays on the order of 30-50% per year as capital crowds in, so a live market eval is self-difficultating, exactly what model comparison needs once benchmarks plateau. The hard part I'd flag is comparability: market paths are stochastic, so two runs of the same model can land on wildly different difficulty depending on the realized path — without controlling for that, the eval measures luck more than the model. Options that work: fixed-seed regime paths with resampled baselines, or bootstrapped difficulty metrics (percentile of PnL against a distribution of random strategies) instead of raw return. The second hard part is reward shaping for RL: sparse PnL rewards over multi-day horizons give terrible exploration, so you'll likely need shaped intermediate rewards (execution quality, information state) to get gradients flowing at all. Also worth publishing: variance across seeds for each model — that's the number that tells users whether a 2% delta is signal or noise. Are you planning to ship fixed seed regimes, or is path stochasticity part of the point?
Extracting feedback from agent conversations is harder than it looks because users don't give explicit feedback — they rephrase their query, which is implicit negative feedback, or they accept the output and move on, which is implicit positive. The signal-to-noise ratio in natural conversation is ~5-10% actionable feedback at best. I'd be interested in how Agnost handles the classification pipeline. Are you using an LLM-as-judge approach (costly but flexible) or a fine-tuned classifier (cheaper but needs retraining)? And how do you handle the cold-start problem where you have no labeled data to train the initial classifier? For enterprise deployments, the integration with existing support workflows (Zendesk, Intercom tickets) is often the deciding factor — can Agnost map agent conversations to existing ticket IDs?
test comment from modgate