· 23 min readFeaturedDeep dive
24 kHz in. 12.5 Hz out.
Inside a speech tokenizer: waveform bandwidth, residual quantization, model context, and streaming delay. Five Manim animations make the arithmetic visible.
- speech
- tokenization
- inference
- latency
Every piece
In build order.
Deep dives carry the code and the failure modes. Practical pieces are the checklists we actually use.
- · 5 min readDeep dive
Speculative decoding in production: when it pays and when it does not
Draft models and n-gram lookahead can cut per-token latency a great deal, or do nothing at all while burning compute. The mechanics, the workloads, and the harness we run before switching it on.
- inference
- vllm
- latency
- · 6 min readDeep dive
A latency budget for voice agents, millisecond by millisecond
A phone call forgives roughly a second of silence before it starts to feel wrong. Here is where those milliseconds go, and how we spend them.
- voice
- latency
- asr
- tts
- · 5 min readDeep dive
Reinforcement learning with verifiable rewards, on one node
You do not need a cluster to teach a small model to follow a schema, call tools correctly and stop guessing. A practical recipe with GRPO, a reward function you can read, and the ways it goes wrong.
- training
- rl
- post-training
- · 5 min readPractical
Evals as a release gate: the harness that runs before every deploy
An agent is a distribution, not a function. The only honest way to ship a change is to measure the distribution before and after. What the harness checks, how it is wired into CI, and what blocks a release.
- evals
- ci
- agents
- · 5 min readPractical
Updating models inside an air-gapped network
No outbound connection means no pulls, no telemetry and no "just download it". How we package, verify and roll a model update through a network that cannot see the internet.
- deployment
- sovereign
- mlops
Short pieces for the people who own an operation: what to automate first, what a voice agent can honestly do, what sovereign really means, and where the money goes.
The blogBook a call
Bring us the hard part.
Inference, speech, post-training, deployment. Twenty minutes is enough to say whether we can take it.
A pilot starts within 5 working days of agreed scope · Nothing upfront · No seat licences