From first call to a permanent AI asset on your infrastructure — in 8 weeks.
Architecture, deployment flow, security model, and the AMC lifecycle. Designed to be readable by the architect who will defend it in an audit.
Layer 1 is what runs on your hardware. Layer 2 is ephemeral cloud compute we tear down after every fine-tuning run. Layer 3 is the artefact — adapter plus eval report — shipped from Layer 2 to Layer 1.
Pre-fine-tuned adapters for the customer's use cases. Versioned in your model registry, swappable per workflow.
Runs in cloud, off-customer-hardware. Spun up per run, torn down after handover. Customer data deleted within 7 days.
The Docker bundle that ships to your server room. Every component is open-source, replaceable, and independently observable.
Layer 1 is what runs on your hardware. Layer 2 is ephemeral cloud compute we tear down after every fine-tuning run. Layer 3 is the artefact (adapter + eval report) shipped from Layer 2 to Layer 1. Each layer is auditable independently.
Each phase declares customer responsibility, Reyatech responsibility, and a deliverable artefact. If a phase doesn't produce its deliverable, the next phase doesn't start.
Provides sample SOPs / batch records / credit memos under NDA. Names champion, auditor, IT contact.
Scopes use cases. Defines eval criteria with customer. Finalises pilot SOW with eval pass-bar.
Signed pilot SOW + eval rubric.
Reviews intermediate eval reports.
Runs LoRA training on cloud GPU. Tears down compute. Deletes training data per §7-day policy.
Adapter file + eval report + certificate of training-data destruction.
Provisions hardware (or accepts shipped bundle). Opens internal firewall to allowed sources.
Deploys Docker bundle. Configures Keycloak with the customer's IdP. Wires Langfuse to the customer's SIEM.
Production system reachable from customer LAN + IQ/OQ artefacts.
Runs user-acceptance testing against the agreed eval rubric.
Triages UAT findings. Runs PQ. Signs go-live cert.
PQ certificate · runbook · handover dossier · AMC start date.
Every component is open-source, replaceable, and independently observable. No closed-source artefacts run in production.
/v1/chat/completions, /v1/embeddings. Routing, fallbacks, rate limits.
All measurements taken on Qwen 3 32B with 8 K context. Capex shown pre-GST in INR.
| Configuration | Users | Model fit | Sustained tok/s | Power (W typ / peak) | Annual electricity (₹ @ ₹8/kWh) | Capex (pre-GST) |
|---|---|---|---|---|---|---|
| Mac Mini M4 Pro 64 GBapple-silicon · inference only | 1–5 | 32B Q4 | 14–18 | 35 / 80 | ~₹3.5 K | ₹2.4 L |
| RTX 4090 Workstation 64 GBsingle-gpu · workstation form-factor | 5–15 | 32B Q4–Q8 | 38–55 | 320 / 600 | ~₹26 K | ₹4.2 L |
| L40S 48 GB Server (1U)rack-mount · datacentre-ready | 15–50 | 32B FP16 | 60–85 | 350 / 700 | ~₹30 K | ₹8 L |
Sustained tok/s measured on Qwen 3 32B with 8 K context — batch 1 for Mac, batch 4 for 4090, batch 8 for L40S. Source: research/Reyatech on-prem LLM TCO.md.
For non-ML evaluators. Three claims, each with the engineering detail behind it.
We use LoRA (low-rank adaptation). The base model stays unchanged; only a small adapter file (~50–200 MB) is trained. Efficient, reversible, auditable.
Training runs on a cloud GPU we rent per-run (currently RunPod A100 80 GB; vendor-substitutable). Your data is encrypted in transit (TLS 1.3), used only for the run, and deleted within 7 days. A signed certificate of destruction is part of every engagement.
Typical fine-tuning cost: ₹500–₹3,000 per training run, paid by us as part of the project fee. You see the eval report from every run — pass-rate per task, regression deltas, sample outputs.
The adapter is yours under §7 of the SOW. The base model is open source. We retain no weights after handover.
Training is ephemeral and cloud-bound. Production is steady-state and on-prem. The two never touch.
Data leaves only on encrypted upload. Cloud volume destroyed within 7 days. Adapter exported back to customer; certificate of destruction signed.
All traffic stays on the customer LAN. Zero outbound calls to the internet are required for inference. Air-gap deployment is supported.
Two columns: included by default, and out-of-scope items that we price separately so there's no ambiguity.
A 30-minute technical deep-dive with the founder. Bring your architecture questions; we'll bring the runbook, IQ/OQ template, and the SOW redlines.