From training to inference, we do the model engineering ourselves
Fine-tuning, reinforcement learning, inference optimisation and a token gateway: one team, deployed privately — or used directly, per token, through one API.
Fine-tuning
Train a model that knows your industry, on your own data
General models don't know your terminology or your internal documents. We clean and label your material into training data, fine-tune an open-weight base model in a private environment, and let an evaluation set built from your real material decide whether it ships.
What it covers
- Data cleaning, de-duplication, de-identification and labelling
- SFT, LoRA / QLoRA and full-parameter fine-tuning
- Evaluation sets from your real material, compared against the base model
- Promoted only when evaluation passes, with one-click rollback
- PyTorchDeepSpeedLLaMA-FactoryLoRA
- Companies with proprietary data that general models get wrong
Reinforcement learning
Use business feedback to teach the model to get the job done
Fine-tuning teaches a model to talk; reinforcement learning teaches it to act. We turn business rules into verifiable rewards so the model tries real tasks, gets scored and updates — aimed squarely at multi-step reasoning and tool-use success.
What it covers
- DPO preference alignment; GRPO / PPO online RL
- Reward models and rule-based verifiers (verifiable rewards)
- Tool-use and multi-step task training for agents
- Production feedback flows back — only after a person confirms it
- TRLverlRayvLLM
- Teams that need agents to finish multi-step business tasks reliably
Inference optimisation
More concurrency and lower latency from the same GPUs
Compute is the biggest cost of a private deployment. We tune layer by layer — quantisation, batching, caching, multi-GPU parallelism — size domestic or consumer-grade GPUs to the use case, and write concurrency and time-to-first-token into acceptance through load tests.
What it covers
- INT4 / FP8 quantisation, with accuracy loss checked on the evaluation set
- Continuous batching, prefix caching and speculative decoding
- Tensor parallelism and multi-GPU deployment with load balancing
- Domestic and consumer-grade GPU sizing, offline deployment
- vLLMSGLangFP8Tensor Parallel
- Companies already running private LLMs and short on concurrency or budget
Token gateway
One API for Chinese LLMs and your private models, metered per token
A unified, OpenAI-compatible model API: change one URL and your systems can route, fail over and bill across multiple Chinese LLM vendors and your privately deployed open-weight models. Every call leaves through one exit, with usage, cost and content fully auditable.
What it covers
- OpenAI-compatible API with per-use-case routing and load balancing
- Automatic failover to a backup model before the first token
- Per-token metering and billing, department quotas and budget circuit-breakers
- Sensitive-data filtering and a full audit log
- OpenAI APIDeepSeekQwenGLM
- Companies and dev teams with many departments or systems calling LLMs
The training line: nothing ships until evaluation passes
One line from data to serving: evaluation-gated promotion, any forced bypass is logged, and every version can roll back in one click.
- 1DataIngest, clean, de-identify
- 2LabellingWith domain experts
- 3TrainingFine-tuning and RL
- 4Evaluation gateNo pass, no promotion
- 5ServeStaged through the gateway
- 6RollbackOne click to the last version
Measured on delivered projects
The figures below come from delivered projects, each with its source and basis: measured in the delivery environment, or the gateway's configured models. They are not long-run production statistics.
9 / 20
Chinese model vendors / models behind one gateway; overseas models disabled at the client's request
−71%
drop in platform p95 tail latency at concurrency 30
56 tok/s
generation speed per stream for a quantised model at 16 concurrent streams
126 mswas 2.7 s
median knowledge-base retrieval time, reranking included
Start with a half-day workshop
Your business owner and IT lead join us to pin down 1–3 candidate use cases and their success metrics.
Book a use-case workshop1
Use-case workshop
- Business owner and IT lead attend
- Map 1–3 candidate use cases and the state of the data
- Output: use-case list and draft success metrics
2
PoC proposal
- Data and system boundaries, evaluation-set design
- Milestones, commitments on both sides, acceptance bar
- Output: a PoC proposal ready to sign
3
Run the PoC
- Your real material, in a private or demo environment
- Weekly check-ins with visible metrics
- Output: evaluation report and launch plan