What I write about
Voice AI
65 articles
Speech recognition, speech synthesis and voice agents: latency, cost and evaluation on real audio.
AI Agents
104 articles
Tool use, memory, multi-agent systems, evaluation and the reliability math of putting agents in production.
AI Security
43 articles
Prompt injection, agent privilege escalation, red teaming and security programs for LLM and voice systems.
ML Systems
80 articles
Serving, inference cost, quantization, KV caches and the build-or-buy decisions behind production ML.
Advisory
2 articles
For founders, CTOs and investors: technical health checks, due diligence and build-or-buy calls.
Research
19 publications, 2 patents
Peer-reviewed work on multilingual speech recognition, speaker personalization and speech synthesis.
Start here
Good first reads if you are new here.
When self-hosted TTS beats the API: the 2026 cost cliff
A worked cost model for when running your own speech synthesis beats paying per character.
Why human-in-the-loop beats full autonomy for high-stakes agent actions
The reliability math behind putting a human approval step into high-stakes agents.
AI security assessments: how to evaluate an LLM system you did not build
How to assess the security of an LLM system you did not build.
ASR is solved (on benchmarks): the real-world gap every voice agent team hits
Why speech recognition benchmark scores overstate how a model will do on your calls.
The build vs buy decision for AI in 2026: a practical framework
A framework for the build-or-buy calls every AI team faces.
How I evaluate AI startups for investors: 8 technical signals
The first-pass technical screen I use before a full due diligence.
Also free to use: the voice agent review checklist, covering latency, recognition, cost, evaluation and security.
Recently published
- SDSL: speculative decoding finally has scaling laws, and they predict your draft model is too big ML Systems
- The build vs buy decision for AI in 2026: a practical framework ML Systems
- Blink: CPU-free LLM inference via SmartNIC, up to 8.47x lower P99 TTFT ML Systems
- The M2M visibility crisis: half of organizations are blind to non-human traffic AI Security
- Jailbreak detection moves inside the model: why text-level filters lost the arms race AI Security
Selected research
Speech recognition and synthesis research, much of it on Indian languages. Google Scholar →
A Unified Approach to Multilingual ASR with Improved Language Identification for Indic Languages
INTERSPEECH 2024: Joint language identification and multilingual recognition on Whisper: 19.1% absolute WER improvement on Indic languages.
Speaker Personalization for ASR using Weight-Decomposed Low-Rank Adaptation
INTERSPEECH 2024: LoRA and DoRA adaptation of a conformer transducer: 20% average relative WER reduction for speakers with little data.
Robust Speaker Personalisation Using Generalized Low-Rank Adaptation for ASR
ICASSP 2024: Generalized LoRA for per-speaker adaptation while keeping compute and memory low.
Robust Speech Recognition Using Meta-Learning for Low-Resource Accents
NCC 2024: Model-agnostic meta-learning for accented English: about 12% relative WER improvement on most low-resource accents.
Working on a related problem? I am open to research collaborations with students, researchers and labs, and to contract research for companies. Let's talk →
Let's talk
If something here is close to what you are working on, I would like to hear about it. I work with companies as a contract engineer, fractional CTO or research partner, run security and architecture reviews, and collaborate with students, researchers and labs. A short email is the best way to start: email me. How I work with teams →