Abhi Ram Salammagari
senior ML engineer · workday · fremont, ca
ML at Workday's HCM platform by day; on-device inference on Apple Silicon after hours. Before Workday I spent five and a half years at [24]7.ai on conversational AI, hold a US patent for agent-training methods, and founded Learnasky in India earlier in my career. I run side projects, contribute to open source, and write down what I learn along the way. This site is where the writing lives.
↳ currently: hunting 3.8 WER points — why identical Parakeet weights score ~6.3% on one runtime and ~2.5% on another
pinned
- Pomvox — fully local, open-source voice dictation for macOS. STT on the Neural Engine, cleanup by SimpleWords v3, my own QLoRA fine-tune, on the GPU.
- Dense Bench: a five-part series on dense 70B inference on M3 Ultra — throughput curves, the Llama 128K cliff, KV-cache quantization, and agentic tool-calling reliability
- Four merged PRs in Google's Workspace CLI (25.1k★) — docs, structured exit codes, input validation, output hygiene
- AI-mmUnity — Google Cloud Next '26 AI Arena hackathon winner
writing
- 2026-08-04 Fine-tuning a dictation cleanup model: designing the corpus, and the category 25% self-correction data still missed.
- 2026-06-24 Dense Bench, Part 4: Memory doesn't predict it — agentic reliability vs quantization.
- 2026-05-12 Dense Bench, Part 3: Compressed KV at the retrieval boundary.
- 2026-05-07 Dense Bench, Part 2: The Llama 128K cliff.
- 2026-05-03 Dense Bench, Part 1: Prefill and decode curves from 4K to 128K context.
elsewhere
press / for command palette