Integrations
Fits the stack you already run
Working together
What your team does, week by week
A typical pilot, from your engineering team's side.
- Week 0
Kickoff and access
Security review, NDA, and a least-privilege role or project in your cloud. Nothing is moved yet.
- Week 1
Data handoff
You share a representative sample in the format below, and we report back on its quality before training starts.
- Weeks 2–4
Training runs
Jobs run on the agreed infrastructure. You get a weekly scorecard and can inspect any checkpoint.
- Week 5
Sign-off
Your domain owner reviews the scorecard and a sample of outputs, and approves the model for release.
- Week 6
Endpoint live
We deploy to the agreed environment and load-test it against your traffic profile before handing over the keys.
- Ongoing
Monitoring
Traces and quality samples go to your observability stack, and any failures they reveal become the next training set.
Base models
Model families we tune
We choose the exact version and size for each project by benchmarking on your data.
| Family | Provider | Access | Methods | Runs in |
|---|---|---|---|---|
| Llama | Meta | Open weights | LoRA, full fine-tune, DPO, continued pretraining | Your cloud, on-prem, edge |
| Mistral / Mixtral | Mistral AI | Open weights | LoRA, full fine-tune, DPO | Your cloud, on-prem |
| Qwen | Alibaba Cloud | Open weights | LoRA, full fine-tune, DPO, continued pretraining | Your cloud, on-prem, edge |
| Gemma | Open weights | LoRA, full fine-tune, distillation target | Your cloud, on-prem, edge | |
| DeepSeek | DeepSeek | Open weights | LoRA, distillation source | Your cloud, on-prem |
| Phi | Microsoft | Open weights | LoRA, full fine-tune | Your cloud, on-prem, edge |
| Hosted frontier models | OpenAI, Google Vertex AI, Amazon Bedrock | Provider API | Supervised fine-tuning where the provider offers it | Provider's managed endpoint |
Infrastructure
Where it trains and where it serves
Clouds and compute
- AWSSageMaker, Bedrock custom models, EKS
- Google CloudVertex AI, GKE
- Microsoft AzureAzure Machine Learning, AKS
- GPU cloudsTogether AI, Fireworks AI, CoreWeave
- On-prem & air-gappedYour own GPUs, no internet egress
Serving runtimes
- vLLMHigh-throughput serving with paged attention
- Hugging Face TGIProduction text-generation server
- NVIDIA Triton + TensorRT-LLMLowest latency on NVIDIA hardware
- llama.cpp / OllamaCPU and edge devices, quantised GGUF
Data handoff
The format we'll ask for
One JSON object per line in the standard chat format. If your data isn't in this shape yet, converting it is part of our work.
Train in place
We run jobs inside your cloud account. Your data never leaves it.
Bucket share
S3, GCS, or Azure Blob, with scoped, expiring credentials.
SFTP drop
An encrypted endpoint that only accepts your IP addresses, deleted when the project closes.
{"messages": [{"role": "system", "content": "You route support tickets for Acme Lending."}, {"role": "user", "content": "EMI debited twice this month, pls check"}, {"role": "assistant", "content": "{\"queue\": \"payments-disputes\", \"reason\": \"DUPLICATE_DEBIT\"}"}]}
{"messages": [{"role": "system", "content": "You route support tickets for Acme Lending."}, {"role": "user", "content": "How do I close my loan early?"}, {"role": "assistant", "content": "{\"queue\": \"loan-servicing\", \"reason\": \"FORECLOSURE_REQUEST\"}"}]}Calling your model
An OpenAI-compatible API
Existing SDKs, frameworks, and tools work as they are: point them at your endpoint and change the model name.
import os
from openai import OpenAI
client = OpenAI(
base_url="https://llm.your-company.internal/v1", # your deployed endpoint
api_key=os.environ["MODEL_API_KEY"],
)
response = client.chat.completions.create(
model="acme-ticket-router-v3",
messages=[{"role": "user", "content": "EMI debited twice this month"}],
temperature=0,
)
print(response.choices[0].message.content)import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://llm.your-company.internal/v1", // your deployed endpoint
apiKey: process.env.MODEL_API_KEY,
});
const response = await client.chat.completions.create({
model: "acme-ticket-router-v3",
messages: [{ role: "user", content: "EMI debited twice this month" }],
temperature: 0,
});
console.log(response.choices[0].message.content);curl https://llm.your-company.internal/v1/chat/completions \
-H "Authorization: Bearer $MODEL_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "acme-ticket-router-v3",
"messages": [{"role": "user", "content": "EMI debited twice this month"}],
"temperature": 0
}'Check your stack
Tell us your cloud, base model, and serving constraints. We'll confirm compatibility.
We reply within one business day (IST).