Generative Flows AIGFAI

Research and development

Applied research with one test: does it make our products measurably better? We work on quality, latency, cost per interaction and safety, and ship what passes.

Focus areas

  • Post-training

    Supervised fine-tuning and preference optimisation to shape behaviour, tone and policy compliance for a specific product.

  • Memory and retrieval

    Long-term memory that keeps conversations consistent over weeks of use, without letting context length and cost grow unchecked.

  • Speech

    Streaming recognition and synthesis with latency low enough for natural back-and-forth conversation.

  • Visual generation

    Diffusion and flow-matching models for image and video, with fine control over style and consistency.

  • Efficiency

    Quantisation, distillation, speculative decoding and caching, to cut latency and cost per interaction without measurable loss in quality.

  • Evaluation and safety

    Model-graded and human evaluation, red-teaming and classifier-based guardrails, run as regression suites on every release.

From hypothesis to production

  1. Hypothesis

    A measurable gap in a product metric, or a published result worth testing on our own workloads.

  2. Prototype

    A small, fast offline experiment, before any engineering investment.

  3. Evaluate

    Offline evaluation and human review against defined quality, safety and cost targets.

  4. Ship

    A staged rollout, with online A/B measurement against the current version.

Research collaborations

We work with universities, research institutes and companies on applied problems in generative AI, from evaluation methods to efficient inference.

Propose a collaboration