I publicly engineer real-world AI systems from scratch — documenting decisions, failures, tradeoffs, and mental models along the way.
Real-time AI voice sales agent. 40+ concurrent calls. Sub-600ms end-to-end latency. Zero vendor lock-in.
| Component | Technology | Latency |
|---|---|---|
| Voice Activity Detection | Silero VAD v5 + WebRTC (dual-fusion) | 6ms |
| Speech Recognition | Faster-Whisper Large-v3 (streaming) | 150ms |
| Language Model | Qwen-2.5-7B-Instruct + KV-cache reuse | 320ms |
| Text-to-Speech | Piper TTS (streaming, sentence-by-sentence) | 100ms |
| End-to-End P95 | ~580ms |
What makes it different:
- 🔁 KV-cache reuse across turns — 40% LLM latency reduction by serializing attention tensors to Redis
- 🎤 Adaptive end-of-turn detection — learns speaking pace per session; 420ms–720ms dynamic silence threshold
- ⚡ Barge-in handling — multi-signal fusion detects interruptions in <200ms, stops TTS mid-sentence
- 🧱 Redis-backed session persistence — full state, KV-cache, metrics per session; horizontal scaling ready
- ☸️ Kubernetes-ready — StatefulSet, HPA auto-scaling, PSTN via Asterisk/FreeSWITCH
Stack: Python FastAPI AsyncIO WebSockets Silero Faster-Whisper PyTorch Redis Docker Kubernetes
→ github.com/dhruvthakur2000/SaleTech
| Project | What it does | Stack |
|---|---|---|
| SaleTech | Production real-time AI voice sales agent — 40+ concurrent calls, <600ms latency, full open-source pipeline | Python · FastAPI · Silero · Faster-Whisper · Qwen · Piper · Redis · K8s |
| linux_driver_eval | CLI framework to benchmark how well LLMs write Linux kernel device driver code. Two pipelines: generation + evaluation. Weighted scoring across correctness, security, quality, performance | Python · GCC · Together API · Static Analysis |
| virtual-voicebot | Streamlit voice assistant with persona-aware responses — the project that started my obsession with real-time audio pipelines | Python · Streamlit · Groq · Whisper · LLaMA · TTS |
| 🔜 HomeAssist (planned) | Smarter Alexa — always-on edge voice assistant using SaleTech's VAD + ASR + buffer layers. Wake-word detection, local LLM, zero cloud dependency | SaleTech core · Edge inference |
| 🔜 SaleTech Analytics (planned) | Call intelligence layer — real-time sentiment, objection detection, sales stage classification per turn | SaleTech core · NLP · Classification |
I write about real engineering decisions — not tutorials copied from docs.
Open to ML Engineer / AI Engineer / Backend Python roles — remote or on-site.
