Daniel is an AI engineer available by subscription. PyTorch and JAX in production, opinions at no extra charge. Pick how much of his attention you want to allocate.
daniel.reply(email) → sampled at temperature 1.2 · no SLA · may hallucinate availability
Testimonials, lightly curated. Selection bias: yes.
He jit-compiled our standup. It's four minutes now.
— CTO, stealth startup (Remote tier)Asked for a quick opinion on our eval suite. Received a working eval suite. Invoice was fair.
— Head of ML, Series B (Async tier)Docked one star because he keeps saying 'it depends.' He is always right, which makes it worse.
— CEO, enterprise (CTO/CAIO tier)Benchmarks self-reported. Evaluation methodology: vibes.
| Spec | Async | Remote | In-Office | CTO / CAIO |
|---|---|---|---|---|
| Attention heads | 1 | 8 | 16 | all of them |
| Context retention | Between threads | Between sprints | Between quarters | Company history |
| Latency SLA | 24h | 4h | shoulder tap | reads your mind* |
| Fine-tuning on your stack | — | ✓ | ✓ | ✓ + distills to your team |
| Human-in-the-loop | Consulted | Is the loop | Is the loop, locally | Designs the loop |
| RLHF | Accepts feedback | Accepts feedback | Accepts feedback, visibly | Gives the feedback |
| Temperature | 0.7 | 0.7 | 0.4 before coffee | deterministic in board rooms |
| Uptime | 99.5% | 99.5% | 99.5% | 99.5% (sleep is a hard dependency) |
Frequently anticipated questions.
Yes — email. It's REST-ish: you send a request, and Daniel returns a response. Rate limits apply. Webhooks are called "calendar invites."
Instantly. Downgrades take one sprint, for emotional reasons.
~2 weeks of onboarding on your codebase and infra. Catastrophic forgetting of your legacy Perl is a feature, not a bug.
A 15-minute chat — grab a slot. Output is watermarked ("as I mentioned on our call…").
Fluent PyTorch, increasingly JAX. Will jit anything that holds still long enough. Functional where practical, pragmatic where not.
The pricing table is real. The costume is the joke. Attention really is the scarce resource — someone proved that in 2017.