About

👋 Hi, I’m Akash, an applied researcher/engineer with experience in speech, audio (at Microsoft), and most recently multi-modal document understanding and retrieval (at Contextual AI). This incidentally completes the trio of audio, vision & text AI multimodality. :)

I’m currently on a sabbatical. After moving to the USA for grad school ~10 years ago, I decided to take a break to reflect, recharge and tinker before setting sail again. More on this here shortly!

Work

Contextual AI

[2024-25]

Wrangled millions of pages to land the first $ millions in enterprise contracts :)

System development (0→1)
- Designed core multimodal document understanding (parsing + ETL index) used by retrieval agents
- Ingesting O(billion) tokens of multimodal documents with complex layouts, tables, graphics, metadata
- Owned system quality. Co-led with platform lead: architected modules to fit multi-stage async distributed pipeline spanning CPU, GPU, and VLM API stages
Applied research: harness and eval design, token-efficient retrieval
- Combining parsing tools, custom segmentation models and VLMs for auditable OCR with bbox provenance and minimal hallucination
- Token-efficient synthesis beyond point queries via ingest-time ETL (à la llm-wiki)
- Blog: Agentic alternative to GraphRAG Demo: Chat with 250 page PDF in Cursor
Tech Lead Manager
- DRI cross-company; Interviewed candidates, Mentored team of 3
- Critical in landing company’s first multi-million $ enterprise contract with Qualcomm

Microsoft

[2018-23]

Fun fact: ~6M hours of monthly traffic equals 1 *year* of conversations transcribed per hour!

Model development: state-of-art transcription designed for scale [O(10 million) hrs/mo]
- Shipped both batch and streaming models to Azure Batch, Microsoft Word, and Microsoft Teams
- Optimized Conformer batch model (Whisper-comparable) at 50x realtime
Applied research: diarized multi-speaker multi-mic transcription
- Shipped diarized in-conference room transcription device covered by The Verge
- Lead contributor: ASR training recipes, evaluation metrics, cross-system error analysis
Research engineering: data pipelines, optimizing distributed training and inference
- Speeding up O(1e20) FLOP training on low-cost V100 GPUs
- Leveraged NVIDIA/ONNX profiling tools to fix bottlenecks in inference throughput
Other Links:
- US Patent US11044287B1: Network resilient real-time voice communication leveraging on-device speech models

‘Graduated’ as one of the few non-speech-PhD senior members on the team :)

For more details, see my resume.

Misc

Open source

[2026] 🎙️ Self-study: efficient audio generation (github, ongoing). Speculative decoding with relaxed acceptance on single-codebook TTS (Chatterbox-TTS); Meanflow distillation for single-step FSQ token flow decoder (Voxtral-TTS).
[2023] 🐥🗣️ Contributed to whisper.cpp (50k+ stars). tinydiarize is a lightweight extension of OpenAI’s Whisper model for speaker diarization, runnable on MacBooks/iPhones.
[2020/26] 🐋 Co-founded OrcaHello, a real-time alert system listening for endangered orca calls 24/7 at underwater “hydrophones” in the Pacific Northwest. Awarded a $30k AI for Earth grant; interviewed in Mongabay News.
[2018] 🗣️ Built Attention, I’m Trying to Speak: speech synthesis with just $75 of compute. Got to fist-bump Richard Socher for Stanford CS224n project award :).

Other

[2025/26] Peer reviewer for ICASSP conference, TMLR journal, and NeurIPS, CVPR workshops.
[2016/17] Wrote case studies on the music streaming industry while studying business/tech strategy at Stanford MS&E.
[2014] Organized (at the time) Chennai’s largest EDM gig - with 5k+ attendees, during my undergrad at IIT Madras/Chennai.