About

A tundish for models.

Tundrel is a production inference and fine-tuning cloud. Independent company. This site describes Tundrel.

About the name

From tundish, the vessel that meters molten metal into a continuous caster. The product meters models onto GPUs the same way: a controlled pour from fine-tune to production inference.

Pronunciation: TUN-drel. The control plane, SKUs, and API hosts use Tundrel names. Domain: tundrel.com.

Team

Rahul

Founder and CEO. Writes the first serving path, signs Caster latency pages, and reads every inbound note.

[email protected]

Pour notes

Mill stencil, short form.

What a tundish is

A tundish meters molten metal into the caster. Continuous. Controlled. Tundrel is named for that pour: models onto GPUs with a rate you can live with. Splash is wasted heat. The product is the meter.

Fine-tune, then pour

A morning LoRA on 50k rows, 3.2 H100 hours, then the same API at lunch on ladle-8b. The checkpoint never left the project. That loop is the company.

Utilization is the foundry metric

Tokens per second look loud. Occupied GPU-seconds pay the hall. Packing, multi-LoRA, and reserved replicas exist so a tray stays hot without stealing a neighbor's p95.

Careers

Four seats on the floor.

Write Rahul at [email protected] with the role in the subject and a note on a GPU or serving system you shipped.

GPU systems engineer

San Francisco or remote US. Own cell images, driver pins, and the H100/A100/L40S hour accounting that hits the invoice.

  • Packing scheduler and replica placement
  • On-call for us-east-1 and us-west-2
  • Compensation: $190k to $240k plus equity

Inference runtime engineer

Kernels, cache, speculative decode, structured out. You publish the lab note or the kernel stays off.

  • fp16 / fp8 / int8 paths per SKU
  • p95 TTFT is a merge gate
  • Compensation: $200k to $250k plus equity

Solutions engineer

Turn a customer checkpoint into a SKU, a quota, and a dashboard the buyer can read.

  • Spout to Ladle migrations
  • Shared Slack on Ladle accounts
  • Compensation: $160k to $200k plus equity

Forward-deployed engineer

Sit on Caster pours. Write the latency page with Rahul. Stay through the first 30 days after cutover.

  • VPC diagrams and private link
  • Travel up to 30% for cutovers
  • Compensation: $180k to $230k plus equity

Security

Fences on the pour.

Encryption

TLS 1.3 in transit. AES-256 at rest for project objects, checkpoints, and JSONL. Customer-managed keys are a Caster add-on at $900 per month.

RBAC

Project roles: owner, operator, viewer. Keys are hashed at rest and scoped to a project. Console sessions expire at 12 hours.

Isolation

Spout uses cell fencing. Ladle reserves a replica. Caster is single-tenant or customer VPC. Training on customer data requires a written job.

Attestation

Controls are mapped to the SOC 2 Trust Services Criteria. Ask Rahul for the current attestation status. Pen-test summaries ship under NDA on Caster.

Contact

Talk to an engineer.

Rahul reads this address. Use the form or write directly.

[email protected]

Legal

Privacy

Model weights, prompts, completions, and training files are customer data. Tundrel processes them to run the job you submitted. We do not train Tundrel foundation models on customer data.

Account fields we store: name, email, company, billing address, and key identifiers. Inbound notes on this form stay in Rahul's mail queue.

Request logs retain 7 days on Spout, 30 days on Ladle, and 90 days on Caster, or a shorter window written on the order. Delete a project and objects are purged within 30 days, backups within 90.

Write [email protected] for an export or deletion request. We reply within 10 business days.