Services
We fine-tune large language models for product teams that have outgrown prompt engineering and RAG. SFT, DPO, ORPO, LoRA, and QLoRA on Llama 3.3, Qwen 2.5, Mistral, Phi-4, and OpenAI/Anthropic custom models. Every engagement ships a versioned eval harness, vLLM or TGI inference with INT4/INT8 quantization, and an MLOps loop that catches regressions before users do. Pilots run 8–12 weeks. Feasibility from 7,500 EUR, full pilot from 38,000 EUR, MLOps retainer from 15,000 EUR/month.
Most AI teams should not fine-tune. Prompt engineering, structured outputs, and retrieval-augmented generation solve 80 percent of production cases at frontier-API cost. The remaining 20 percent is where fine-tuning earns its keep — behavioural change that no prompt can force, structured-output adherence above 99 percent, latency under 400 ms at p95, or an inference cost curve that breaks at scale. We help teams figure out which side of that line they are on, then ship the smallest model that hits the eval bar. Engagements start with a written feasibility memo and a versioned eval harness; nothing trains until both are agreed in writing.
Golden-set construction, labelling rubrics, inter-annotator agreement tracking, PII redaction with Presidio, synthetic data generation with frontier models, and deduplication. The labelling pipeline lives in your cloud, not ours.
Supervised fine-tuning for behavioural change, DPO and ORPO for preference alignment without a separate reward model, and KTO when preference data is asymmetric. TRL, Unsloth, and Axolotl on your cloud GPUs or ours.
Parameter-efficient fine-tuning cuts GPU memory 60–80 percent and lets you iterate ablations on a single A100. We default to QLoRA 4-bit for first-pass, full fine-tune only when ablations prove the lift is real.
Versioned golden set, LLM-as-judge with rubric scoring (Ragas, custom), task metrics (F1, BLEU, ROUGE, exact match), adversarial probes for hallucination, jailbreak, PII leakage. CI blocks any merge that regresses the bar.
Training-data versioning with DVC or LakeFS, experiment tracking on Weights & Biases or MLflow, scheduled retraining triggered by drift metrics, and rollback runbooks. Everything reproducible, everything in your repo.
vLLM, TGI, or TensorRT-LLM with continuous batching, INT8/INT4 quantization (AWQ, GPTQ, FP8), speculative decoding, and prefix caching. Load-tested at your real p95 traffic before cutover, with monitoring dashboards.
Week 1–2: written memo answering whether fine-tuning is the right tool, which base model fits the task, expected eval lift over RAG/prompt baseline, and total cost projection over 12 months. Go/no-go before any GPU spend.
Weeks 3–5: golden set of 200–1,000 labelled examples, labelling rubric, PII redaction pipeline, eval harness wired to W&B with frontier baselines. Nothing trains until the eval suite runs green against the baseline.
Weeks 6–8: SFT, then DPO or ORPO if preference data exists. LoRA/QLoRA first, ablations on rank, learning rate, and base model. Every run is reproducible from the config file in your repo.
Weeks 9–12: vLLM/TGI deployment, quantization, load tests at p95 traffic, canary rollout with rollback runbook, monitoring dashboards, and engineer handover. Optional MLOps retainer for continuous training.
Two-week written memo: base-model recommendation, expected eval lift vs RAG/prompt baseline, GPU cost projection, total cost of ownership over 12 months, go/no-go decision. Credit applied to pilot if you proceed. 7,500 EUR fixed.
8–12 weeks. One production model, full eval harness, vLLM/TGI inference deployment in your cloud, load-tested rollout, monitoring dashboards, and engineer handover. Includes 30 days post-launch support. 38,000 EUR fixed.
Continuous training, eval expansion, drift detection, monthly model refresh, vendor cost optimization, on-call for inference incidents. One senior MLE plus eval support, six-month minimum. From 15,000 EUR/month.
All pricing excludes GPU compute cost — we work on your cloud account and you pay AWS/GCP/Azure directly. Typical pilot GPU spend is 2,500–8,000 EUR.
Native iOS and Android e-signature clients with a Symfony + React CRM for a cross-border law firm — KYC onboarding and a defensible evidence trail for US & EU matters.
Production social platform — App Store + Google Play, live across the US and EU — with geo Radar, encrypted messaging and a virtual economy.
B2B e-commerce and product configurator for a global polymer manufacturer with multi-region pricing, stock and dealer workflows.
GDPR-aligned · ISO 27001 ready · SOC 2 Type II in progress · HIPAA-capable · CCPA-acknowledged
No training run starts until the eval harness runs green against your frontier baseline. Every release ships with a regression report. If the eval bar slips, the merge is blocked — not negotiated.
The MLEs on your engagement have shipped fine-tuned models to production. They know what LoRA rank to pick, why your DPO loss diverged, and how to debug a vLLM throughput cliff — without a Twitter thread.
Training runs in your VPC, weights stay in your S3/GCS, code lives in your repo. We work via assumed IAM roles. No data ever lands on our laptops, and you own the resulting model on day one.
For regulated workloads we sign HIPAA BAAs, run on HIPAA-eligible regions only, and integrate with your existing data governance — not parallel to it.
Three signals justify fine-tuning. First, latency or cost: a fine-tuned 7B model on vLLM costs roughly 1/40th of GPT-4o for the same task at p95 latency under 400 ms. Second, behaviour you cannot prompt your way into: domain-specific style, structured output adherence above 99%, or refusal patterns that frontier safety layers will not allow. Third, data leverage: you have 5,000+ high-quality labelled pairs that nobody else has. If the answer is mainly knowledge retrieval, do RAG first. If it is occasional formatting, prompt-engineer first. Fine-tuning is the right call when you need behavioural change at scale.
Both. OpenAI fine-tuning (GPT-4o, GPT-4.1, GPT-4o-mini), Anthropic via Bedrock custom models, Google Gemini tuning, and the full open-source stack: Llama 3.3, Qwen 2.5, Mistral, Phi-4, DeepSeek. The choice is engineering, not ideology. Closed models give you faster delivery and zero infra. Open models give you ownership, cheaper inference at scale, and on-premise deployment when compliance requires it. We run the same eval harness against both paths and present the cost-per-token, latency, and quality trade-off in writing before you commit.
Every engagement ships with a versioned eval suite: a golden set of 200 to 1,000 labelled examples curated with the client, automated LLM-as-judge with rubric scoring (Ragas, custom rubrics), task-specific metrics (BLEU, ROUGE, exact match, F1, structured-output adherence), and adversarial probes for hallucination, jailbreak, and PII leakage. Every training run posts to Weights & Biases with the full eval table. CI blocks any merge that regresses the golden set by more than the agreed threshold (typically 2%). The eval suite is yours, version-controlled in your repo, and runs against frontier baselines on every release.
Parameter-efficient methods first: LoRA and QLoRA cut GPU memory by 60 to 80 percent and let us run a Llama 3.3 70B SFT on a single A100 80GB node for under 300 EUR. Unsloth and Axolotl give us 2x training throughput vs vanilla Hugging Face TRL. We default to QLoRA 4-bit for first-pass experimentation, switch to full fine-tune only when ablations prove it moves the eval needle. Inference cost is controlled by INT8/INT4 quantization (AWQ, GPTQ), vLLM continuous batching, and speculative decoding. A typical client moves from 18,000 EUR/month frontier API spend to 3,500 EUR/month self-hosted inference.
Engagement starts with a mutual NDA and a GDPR-aligned DPA. Training data lives in your cloud account: we work via assumed IAM roles, never copy data to laptops, and the training cluster runs in your VPC (AWS SageMaker, GCP Vertex, Azure ML, or your Kubernetes). For regulated data we sign HIPAA BAAs and run on HIPAA-eligible regions only. PII redaction pipelines (Presidio, custom regex + NER) are part of the data curation step. We are GDPR-aligned, ISO 27001 ready, SOC 2 Type II in progress, HIPAA-capable, and CCPA-acknowledged.
Eight to twelve weeks for a first production model. Weeks 1 to 2: feasibility and eval harness design. Weeks 3 to 5: data curation, labelling pipeline, and golden set construction. Weeks 6 to 8: SFT plus DPO/ORPO training runs, ablations, and eval iteration. Weeks 9 to 10: inference serving (vLLM or TGI), quantization, load testing. Weeks 11 to 12: canary rollout, monitoring dashboards, runbooks, and handover to your team. After that we either move to a retainer (continuous training, eval expansion, drift response) or step out cleanly with documentation.
