Your model works in a notebook. We get it running on GPUs, in your cloud.
TrigunAI deploys open models, RAG and GPU pipelines on NVIDIA hardware for startups worldwide. You get a working endpoint in your own AWS, GCP or Azure account, measured for speed and cost, with the code and a runbook your team can keep.
- Model
- 8B instruct, FP8
- Serving
- vLLM, OpenAI API
- GPU
- 1 × L4 24 GB
- p95 latency, 1st token
- 410 ms
- Throughput @ 16 users
- 620 tok/s
- Cost per 1M tokens
- $0.38
Six jobs we take from prototype to production
Self-hosted model APIs
Open models (Llama, Qwen, Mistral, Nemotron and others) behind a private, OpenAI-compatible endpoint, so you stop paying per token and keep data in your account.
vLLM · NVIDIA NIM · Docker · K8s or serverless GPU
Search and answers over your data
Ingestion, chunking, embeddings, a vector store and evaluation, so you know the answer quality before your users do.
pgvector · FastAPI · eval sets · rerankers
GPU image and video pipelines
Diffusion, avatar and lip-sync models wrapped in a job queue that renders reliably and shuts the GPU down when idle.
PyTorch · SDXL · AnimateDiff · MuseTalk · async jobs
Real-time voice agents
Speech-to-text, LLM and text-to-speech wired for low latency, on the phone or the web, with your tools and data.
streaming STT/TTS · LiveKit · telephony · RAG
Simulation and robot learning
Isaac Sim and Isaac Lab environments, RL training runs and ONNX export for on-device inference.
Isaac Sim · Isaac Lab · PyTorch · ONNX
GPU cost and speed audit
Right-size the GPU, batch and quantise, cache, and scale to zero. We measure before and after so the saving is a number, not a promise.
profiling · quantisation · load tests
We run GPU systems for our own products
These are ours, so you can check them before you talk to us.
Autonomous AI video agent
Writes a script with an LLM, voices it, lip-syncs an avatar, renders on a cloud NVIDIA A10G GPU and publishes to YouTube, Instagram and Facebook without a human step. 80+ videos published.
youtube.com/@TrigunAI-InnovationsAcharya AI tutor
An agentic LLM tutor in production: FastAPI, Postgres and Azure Container Apps, with auth, admin and billing.
acharya.trigunai.comMulti-model video generation API
SDXL, AnimateDiff, MuseTalk and SadTalker behind one FastAPI service with async GPU job handling.
github.com/absolutedimension/TrigunVideoForgeFranka peg-in-hole in Isaac Lab
Contact-rich robot manipulation with force feedback: 100% success at 1 mm clearance across seeded trials, with a noise-sweep evaluation.
github.com/avinash246813579/franka-peg-in-holeMore of our work, including XR apps on the Meta Store and voice agents: build.trigunai.com.
From brief to handover, with no surprises
Send a brief
5 minutes. We reply within 1 working day.
Free 30-min call
We ask about your model, users, data and budget, and tell you plainly if we are the wrong fit.
Fixed quote
A written scope with acceptance targets: latency, throughput, cost, quality. Or a paid audit first if the scope is unclear.
Build in your account
We work in your cloud with a scoped role. Short written update every working day.
Handover
Measured report, infrastructure as code, runbook and a recorded walkthrough for your team.
Fixed packages, or pay by the hour
GPU Deployment Audit
3 working days · fixed price
- Review of your model, traffic and current setup
- Recommended GPU, serving stack and cloud
- Cost per request estimate and a build plan
- Written report you can use with any team
Model API Launch
About 1–2 weeks · one model, one cloud
- Open model served on an NVIDIA GPU in your account
- OpenAI-compatible API with keys and rate limits
- Autoscaling, logs, dashboards and alerts
- Load test report, Terraform and runbook
Production Pipeline
About 2–4 weeks · RAG, voice or media
- Everything in Model API Launch
- RAG with evaluation, a voice pipeline or a GPU render queue
- Integration with your app and data
- Two weeks of support after go-live
GPU, cloud and licence costs (for example NVIDIA AI Enterprise for NIM in production) are billed by your provider to your account, not by us. Final price depends on scope. See engagement terms.
Your accounts, your code, your data
Where we are: Patna, India (IST, UTC+5:30). We overlap 3–4 hours with US East Coast mornings and most of the European working day, and we reply to messages within one working day.
What we have run in production, and what we deploy from NVIDIA's recipes
Running in our own products
PyTorch on NVIDIA GPUs (A10G, RTX) · CUDA containers · diffusion and lip-sync models · async GPU job queues · Isaac Sim / Isaac Lab · ONNX export · FastAPI · Postgres · Docker · AWS and Azure
Deployed from NVIDIA's documented recipes
NVIDIA NIM · TensorRT-LLM · Triton Inference Server · Riva speech · DeepStream · Jetson
If your project depends on a tool from the second list, we will tell you on the first call how much of it we have done before, and we will not bill you for our learning time. We would rather lose a project than overpromise on one.
Two senior engineers, no account managers
Deepak Kumar · Founder
18+ years in software, including banking systems at TCS, Mastek and 3i Infotech. Builds and runs TrigunAI's LLM, voice and GPU media systems in production.
Avinash · Co-founder
Robotics and XR engineer. Builds the Isaac Sim / Isaac Lab work and shipped TrigunAI's Meta Quest apps.
You contract with TrigunAI Innovations Private Limited, a registered Indian company and a member of NVIDIA Inception.
Common questions
Are you part of NVIDIA?
No. We are an independent company and a member of NVIDIA Inception, NVIDIA's program for startups. NVIDIA does not provide, endorse or guarantee our services.
Who pays for the GPUs?
You do, directly to your cloud or GPU provider, so you always see the real cost. We help you pick the cheapest GPU that meets your targets and set it to scale down when idle.
Do we need NVIDIA AI Enterprise?
Not always. vLLM and many open models are free to run. NVIDIA NIM is free for development, but production use needs an NVIDIA AI Enterprise licence. We explain the trade-off in the audit.
What if you don't hit the targets?
Fixed packages have written acceptance targets agreed before we start. If a target turns out not to be reachable on your budget, we tell you as soon as we know and agree a new plan with you before spending more of your money.
Can we hire you through Upwork?
Yes. If you prefer Upwork's escrow and contracts, say so in the form and we will send you our profile link. If we first met you on Upwork, we keep all work and payments on Upwork.
How do we pay?
We invoice in US dollars from India. You can pay by international bank transfer. Fixed packages are paid in milestones, and hourly work is billed weekly.
Can you work with our in-house team?
Yes. Many teams want a working setup and then to own it. We pair with your engineers, write the runbook for them and record a walkthrough.
Tell us what you want to deploy
No payment and no commitment. We reply within 1 working day with questions or a time for a call.
Prefer email? deepak@trigunai.com
Please do not include passwords, API keys or confidential data in this form. We can sign an NDA before you share details.
Brief received
Thank you. We will reply by email within 1 working day. Your reference is .
You had sent a brief for this company before, so we updated it.
Anything to add? Email deepak@trigunai.com and quote your reference.