Machine Learning Practitioner
Building: Helpish — meeting intelligence, solo-built, production-deployed.
Helpish · AI Assistant for tech product demos
Layered async pipeline — audio capture → transcription → analysis → persistence, each stage swappable.
Browser recorder with mic + tab audio mixing, live transcription via Deepgram, speaker diarisation, voice profile embeddings.
Live sales coaching, cross-session memory with contradiction detection, post-call summaries that adapt per user.
Chronicle · Self-hostable wearable AI backend · 60+ ⭐ · Active contributor
- Added Google Drive, Dropbox, and WebSocket as first-class audio sources; co-designed plugin architecture for third-party apps
- Built speaker diarisation + identification benchmark suite — measures accuracy, latency, speaker confusion across diverse recordings
- Implementing MCP connector support enabling plugins to interact with external services via natural language tool calls
I implement papers to actually understand them. Each repo has unit tests validating against reference implementations.
| Repo | What it covers |
|---|---|
| Stable_Diffusion_from_scratch | Full diffusion architecture in PyTorch |
| Transformers_from_scratch | Attention, positional encoding, encoder-decoder |
| LoRA-implementation | Low-rank adaptation from paper to training loop |
| BERT_from_scratch | Masked LM + NSP pretraining |
| Swin_Transformer_from_scratch | Shifted window attention |
| GPU_programming | CUDA kernels, memory hierarchy, parallel patterns |
| SonicSpeech | Deep learning architectures for speech and audio |
| Dillusion | SOTA diffusion paper implementations |
| Backprop | Scalar autograd from scratch → micrograd |
Cross-lingual embedding alignment (English ↔ Hindi/Gujarati): Procrustes alignment on FastText embeddings — word translation and multilingual similarity search in a shared vector space.
ML & Training: PyTorch, LoRA, SFT, fine-tuning, RAG, GraphRAG, agentic systems, embeddings, eval pipelines
Speech & Audio: Deepgram, Whisper, VAD, speaker diarisation, voice-profile embeddings
Infra: FastAPI, Redis, Docker, AWS, queue-based async architectures
Low-level: CUDA, C/C++, parallel programming
- Hackathon winner ×10
- IEEE: APBTMS — computer vision for real-time bus passenger tracking and overcrowding detection
- Cohere AI Gujarati LLM — top contributor to dataset creation and curation for a low-resource Indian language model
- Currently exploring: concurrency-chaos — threads, async, queues, Redis, event-driven systems
- Reading: Goodreads
- Hack demos: YouTube