# ClutchCall

> Real-time infrastructure for physical and multimodal AI. One QUIC substrate. Every modality — voice, robotics, multiplayer netcode, browser — on the same transport line.

ClutchCall is a bare-metal C++ QUIC transport substrate. Modules
compose verticals; the substrate stays the same. Realtime voice AI runs on
the same line as inference streaming, WebRTC, SIP telephony, MQTT/ROS2
robotics, and games — all multiplexed on one connection, no head-of-line
blocking between modalities.

## Substrate

- Bare-metal C++23, purpose-built for QUIC — no per-modality media servers to run
- 36,732 concurrent MoQT-over-WebTransport connections on one 8-vCPU / 8 GB GCE box at about 0.4 % CPU (0.03-0.04 of 8 cores), no media flowing
- 122,367 messages/second on one 8-core box — improving
- 24,000 concurrent G.711 calls on one GCE c4-highcpu-8 (8 vCPU, 16 GB), every packet delivered, at 52 % CPU
- 1,500 live 1080p viewers on one GCE c4-highcpu-8, nothing lost, 9 Gbps out at 61 % CPU

## Inference (measured on one GPU: RTX A5000 at $0.27/hr, or H100 where named)

- Text to speech, time to first audio p50 / p95: Qwen3-TTS 1.7B 53 / 76 ms (same model on the Coval board: Nari Labs 63 / 91, Baseten 104 / 132)
- Speech to text, end of speech to final transcript p50 / p95, streaming route: Parakeet TDT 0.6B v3 42 / 62 ms on one H100, 49 / 75 ms on one RTX A5000 (same model on the Coval board: Together 49 / 139); Whisper large-v3 64 / 122 ms on one H100, 103 / 206 ms on one RTX A5000 (Baseten 112 / 248, Together 108 / 232); Parakeet TDT 0.6B v2 36 / 52 ms on one H100, 46 / 77 ms on one RTX A5000
- Speech to text accuracy: Qwen3-ASR 1.7B word error rate 3.67 % over the seven English test sets that Coval uses (our run)
- These latencies are our own runs of Coval's open benchmark method, not Coval leaderboard entries. Board values were read on 2026-10-14. STT: our seed-0 sample of the board's 7 sets x 50 clips, client about 10 ms from the GPU. TTS: Coval's public prompt set; the board uses a private one
- 64 concurrent Qwen3-TTS calls per card; 64 live Qwen3-ASR streams with partial transcripts per card
- Hindi speech to text: IndicConformer 600M 76 / 103 ms, WER 13.1 % on FLEURS
- Voxtral-4B-TTS at 64 calls on the same card: first audio ~60 ms on our engine vs >6 s on vLLM-Omni
- Robot policy: π0.5 (DROID) about 27 ms model time (p50 26.6 ms, p99.9 26.8 ms) per action chunk on an RTX PRO 6000 with 8-bit (W8A16) weights, on our C++ runtime inside our serving stack

## Robot control over our relay (our own runs, 2026-10-17 and 2026-10-18)

- Relay round trip, robot -> relay -> responder -> relay -> robot, 40 KB observation at 50 per second, a stand-in answers in place of the model: p50 4.6 ms, p99.9 7.7 ms, 20,000 round trips per run, 0 failures (production relay in Mumbai; robot and stand-in in the same metro)
- Across cities, robot and GPU host in Noida over our relay in Mumbai, 56 KB π0.5 observation (3 camera images + prompt), a stand-in answers in place of the model: p50 46 ms, p99.9 64 ms, 20,000 round trips, 0 failures
- π0.5 control step on the same Noida -> Mumbai path, camera frames in and action chunk out, π0.5 with 8-bit (W8A16) weights on one RTX PRO 6000: p50 71 ms (model 26.6 ms, network 44.5 ms), p99.9 82 ms, 20,000 steps, 0 failures, none over 200 ms
- Every round trip is timed on the robot by its own clock; a request with no answer counts as a failure

## Use cases

- [Voice AI](https://clutchcall.dev/voice): From the phone in a hand to the model and back — supervised, the whole way on one line.
- [Inference](https://clutchcall.dev/inference): Tokens straight off the GPU to the client you already use — and the GPU stops the moment they stop mattering.
- [Interactive live](https://clutchcall.dev/streaming): Sub-second live streams for classes, live selling and watch-alongs, where 6 seconds of HLS delay breaks the conversation.
- [Browser · WebRTC](https://clutchcall.dev/webrtc): Every modality shares a single line, and none can block another.
- [SIP · Telephony](https://clutchcall.dev/sip): Connect the way carriers expect. Then get out of the media path.
- [Robotics](https://clutchcall.dev/robotics): Your fleet stays in control on-site and visible from anywhere.
- [Games · Netcode](https://clutchcall.dev/games): Datagrams and reliable channels on one session, with a path through firewalls on port 443.

## Key surfaces

- [Homepage](https://clutchcall.dev/): the full narrative — one packet becomes one substrate
- [Docs](https://docs.clutchcall.dev): API references, SDK documentation, architecture guides
- [GitHub](https://github.com/clutchcall): SDKs in Python, TypeScript, C++ and Rust
- [Status](https://clutchcall.instatus.com): live system status

## Pricing

Every engagement starts with a paid 6-week pilot on real traffic,
with written success criteria; the fee is credited to the first
month. Production is priced per concurrent channel per month,
with telephony passed through. On-prem is licensed per node.
Email engineers@clutchcall.dev to start.

## Legal

- [Privacy](https://clutchcall.dev/legal/privacy)
- [Terms](https://clutchcall.dev/legal/terms)
- [Cookie Policy](https://clutchcall.dev/legal/cookies)

## Contact

- engineers — support@clutchcall.dev
- privacy — privacy@clutchcall.dev
- security — security@clutchcall.dev

Operated by Throughgen Innovations Private Limited.