Lead the AI & real-time stack of two AI interviewers at InCruiter: 1,000+ concurrent interviews in 42+ languages on LiveKit Agents, Vapi telephony and Azure OpenAI, with feature-flagged releases to LiveKit Cloud. |
13 merged PRs in other people's repos, 7 of them in LiveKit (agents, python-sdks). Every one starts with a test that fails on |
7 papers rebuilt from scratch in PyTorch, each with a check script: Transformer, GPT + BPE, Whisper, Conformer + RNN-T, LLM.int8/GPTQ/FlashAttention, LoRA/DPO/GRPO and speculative decoding. |
Looking for: Voice AI Engineer · Agents / Applied AI Engineer · Real-time ML · Forward Deployed Engineer
Based in: India (IST) · remote-first, comfortable overlapping with US and EU teams
📫 hassankazinasik@gmail.com · resume on request
| Project | Merged fix | Why it mattered |
|---|---|---|
| LiveKit Agents | #7741 · #7742 | the call recorder no longer hangs when its encoder fails · the TTS markup tokenizer survives stray closing tags |
| LiveKit python-sdks | #860 · #861 · #862 · #864 · #865 | webhook sha256 check · Twirp base path behind proxies · AudioMixer survives a slow input · correct I420A/I422/NV12 plane layout · canSubscribeMetrics grant |
| Piper TTS (Open Home Foundation) | #328 | Hebrew voices get the Hebrew phonemizer in PyTorch inference |
| Mastra | #27102 | Sarvam listen() uses the configured model; three more of my bug reports fixed by the team |
| RealtimeSTT / RealtimeTTS | #329 · #330 · #383 | Hindi/Bengali transcripts keep their final full stop · feed_audio on tiny chunks · ElevenLabs voice settings sent |
| stream2sentence | #16 | Urdu sentences split at ۔ |
All merged PRs · In review now (LiveKit, Pipecat, NVIDIA NeMo, sherpa-onnx, Piper)
|
I build real-time voice agents, and I keep asking one question: why do voice agents break on real Indian speech, and how do we fix it?
|
|
| Paper | Repo | What it shows | |
|---|---|---|---|
| 🧠 | Transformer (Vaswani 2017) | attention-is-all-you-need-hindi | English → Hindi on 1.36M sentence pairs |
| 🧠 | GPT + BPE | gpt-from-scratch | 12M-param GPT with its own tokenizer |
| 👂 | Whisper (Radford 2022) | whisper-from-scratch | exact audio front-end, WER/CER on English and Hindi |
| 👂 | Conformer + RNN-T | conformer-rnnt | streaming ASR with transducer loss and beam search |
| ⚡ | LLM.int8 · GPTQ · FlashAttention | quant-flash | measured perplexity on Qwen2.5-0.5B, Triton kernel |
| 🎯 | LoRA · DPO · GRPO | post-training | preference training that makes a model answer Hindi in Hindi |
| 🚀 | Speculative decoding | speculative-decoding | Leviathan et al. 2023, Algorithm 1 on a hand-written Llama |
|
|
Building voice AI for your users? Let's talk.









