PhiGrad AI logoPhiGrad AIAI engineering for production

LLM & SLM engineering

Custom LLM development, fine-tuning and RAG

We adapt models to your domain with retrieval systems, fine-tuning and alignment, synthetic training data and honest evaluation, so quality is something you can measure instead of hope for.

Typical starting points

  • A RAG system that retrieves the wrong context
  • Prompts stretched far past what prompts can do
  • Too little labelled data to fine-tune on
  • No evaluation set, so quality is anecdotal
  • A frontier model too expensive for the volume
Discuss your model project

Capabilities

What we do at the model layer

  1. 01

    RAG development services

    Retrieval that actually retrieves: chunking and indexing tuned to your documents, hybrid search, reranking, grounded answers with citations and measured answer quality.

  2. 02

    Custom LLM development

    Domain-specific models built on open-weight bases, adapted to your terminology, formats and task structure instead of prompt-patched into approximate behaviour.

  3. 03

    Fine-tuning and alignment

    Supervised fine-tuning, LoRA/QLoRA and other PEFT methods, plus preference alignment with DPO, ORPO, KTO or RLHF where behaviour matters more than raw accuracy.

  4. 04

    Synthetic data generation

    When labelled data is thin, we generate and validate task-specific training data, seeded from your real examples, deduplicated, quality-filtered and checked for leakage.

  5. 05

    Evaluation harnesses

    Task-level benchmarks built from your own data, so every fine-tune, prompt change or model swap is judged on the work it has to do.

  6. 06

    Smaller models, lower cost

    Distilling a frontier-model workflow into a small model you can serve cheaply, with quantization, serving setup and latency work included.

Choosing an approach

Retrieval, fine-tuning, or both

The decision should come from measurements, not fashion. We set up the evaluation first so the choice is evidence-based.

When retrieval is the right answer

Knowledge changes often, answers must cite sources, and the model needs facts rather than new behaviour. Usually faster, cheaper and easier to keep current.

  • Knowledge assistants over internal documents
  • Support answers with citations
  • Frequently changing policies or catalogues
  • Strict traceability requirements

When fine-tuning is the right answer

The task has a fixed shape, a house style or a format the base model keeps missing. Or the volume makes a smaller specialised model far cheaper to run.

  • Structured extraction and classification at volume
  • Domain language and tone the base model gets wrong
  • Latency or cost targets a large model cannot meet
  • Offline or private deployment requirements
Explore all servicesAI agent developmentBook an AI architecture consultation

Questions

Common questions

Should we fine-tune or use retrieval?
Often both, but rarely at once. We usually get a retrieval baseline and an evaluation set working first, then fine-tune only where the measurements show it pays: behaviour, format, latency or cost.
We do not have enough labelled data. Is fine-tuning off the table?
No. Synthetic data generation seeded from a small set of real examples covers many tasks, provided it is validated and held to a real evaluation set rather than trusted on faith.
Do you use open-weight or hosted models?
Whichever fits the constraints. Hosted models are often right early; open-weight models win when you need private deployment, per-task cost control or a model you can genuinely own.