Generative AI
Most AI prototypes never survive contact with real users, real data, and real operational scrutiny. Our practice exists to bridge that gap — building the unglamorous infrastructure that turns a clever notebook into a system your team can actually rely on.
Generative AI development is the engineering discipline of building production systems on large language models — retrieval-augmented generation, agentic workflows, and LLM applications — with the retrieval, evaluation, and guardrail infrastructure that keeps them accurate under real load. TrueLeaf Tech builds these model-agnostically across Claude, GPT, Gemini, and open-weight models, so clients are never locked to one provider.
Our AI practice
We've been working with large language models since the first GPT-3 betas. The lessons we've drawn show up in every engagement.
Almost every useful AI feature lives or dies on what gets retrieved before the model sees the prompt. We invest heavily in chunking strategies, hybrid search, and the evaluation harnesses that tell you when retrieval is silently failing.
We build agents with explicit state, narrow tool surfaces, and human-in-the-loop checkpoints where they matter. Our preference is for agents that do one thing dependably, rather than ones that try to do everything and occasionally embarrass themselves.
If you can't measure it, you can't ship it. Every AI feature we build comes with a test set, an evaluation harness, and a clear answer to the question "how do we know this didn't get worse this week?"
What we build
Internal copilots grounded in your documents, with audit trails and role-based scoping built in.
Agents that automate well-defined operational workflows — with state, retries, and human approval steps.
Extraction, classification, and structured output pipelines that beat traditional OCR-plus-rules architectures.
Chat surfaces with memory, tool use, and the safety scaffolding to deploy them in customer-facing contexts.
Custom eval harnesses that catch regressions, measure quality, and let your team ship AI changes confidently.
Orchestration layers that pick the right model for each request, balancing cost, latency, and quality.
Retrieval and generation pipelines that handle non-English content as a first-class concern, not an afterthought.
Input/output guardrails, prompt injection defense, and policy enforcement that holds up in adversarial settings.
How we engage
AI projects fail in predictable places — an unclear success measure, retrieval nobody tested, a pilot that can't be operated. Our sequence is designed to surface those early, while changing course is still cheap.
Before any model is chosen, we agree what "working" means and how it will be scored. A retrieval system with no accuracy target is a research project, not a deliverable. This is usually a week, and it is the week that decides whether the rest succeeds.
We assemble a test set from your real data and wire it into CI before feature work starts. It means every subsequent change is measured rather than argued about, and it is what lets you upgrade models later without fear.
One workflow, one user group, real production data. A narrow slice in front of real users teaches more in two weeks than a broad prototype teaches in two months — and it gives you something to cancel cheaply if the economics don't hold.
Observability, guardrails, cost controls, runbooks, and the documentation your team needs to own it. We would rather write ourselves out of a dependency than sell an operating contract you didn't ask for.
Selected work
The clearest evidence of how we work is what we have already shipped.
A pricing and competitive-intelligence platform built on a multi-million-product corpus, with field-aware delta ingestion keeping data current without reprocessing everything nightly.
Content quality and enrichment tooling operating at marketplace scale, where the cost of a wrong automated edit is measured across millions of listings.
Extraction and classification pipelines that replaced an OCR-plus-rules architecture, with an evaluation harness proving the change was an improvement rather than a lateral move.
An analytics product turning raw marketplace signal into decisions sellers act on, built to stay affordable as the underlying data volume grew.
A generative video pipeline with per-clip cost modelling, model routing, and the credit economics to make an expensive workload commercially viable.
Operational tooling in an environment where reliability and auditability are not negotiable, and where a confident wrong answer is worse than no answer.
Field notes
Engineering guides from the work above — the architecture decisions, trade-offs, and failure modes we hit in production. Written for engineers evaluating this seriously, not for search engines.
LangGraph, CrewAI, AutoGen and the rest — what each is actually good at, and when a framework costs you more than it saves.
The five components every agent shares, single-agent versus multi-agent, and the patterns that survive production.
When adding an agent to retrieval genuinely improves answers, and when it just adds latency and cost.
Hybrid search, reranking, query rewriting, and chunking strategies — with the evaluation setup that tells you which actually helped.
A ground-up walkthrough of the tool loop, state handling, and error recovery, without a framework in the way.
Wiring an agent into a React front end: streaming, tool-call rendering, and interruption handling.
Where agentic automation fits on a plant floor, and the constraints that rule most of it out.
Dispatch, routing, and exception handling — drawn from building last-mile dispatch systems.
Treating non-English content as a first-class concern rather than a translation step bolted on at the end.
Why we hand over the test suite as the primary artefact, and what changes when a client owns their own quality bar.
The review checklist we run against any RAG system, including the failure modes that stay silent until they're expensive.
The case for narrow tool surfaces, explicit state, and resisting the newest abstraction.
Common questions
It builds production systems on top of language models: retrieval pipelines that ground answers in your data, agents that carry out multi-step work, evaluation harnesses that measure quality, and the guardrails and observability needed to run all of it safely. The model itself is rarely the hard part — the surrounding engineering is.
Whichever fits the workload, and we keep the choice reversible. We route per task on cost, latency, and measured accuracy, behind an abstraction that lets you switch providers without rewriting the application. Being locked to one vendor's roadmap is a business risk, not an architecture.
A narrow production slice typically runs six to ten weeks: roughly a week to agree the success measure, two to build the evaluation harness and retrieval layer, and the remainder to ship and harden one real workflow. Broader programmes are sequenced as several such slices rather than one long build.
Usually not as a precondition. Retrieval quality depends far more on chunking, metadata, and query strategy than on which store you pick, and picking the store first tends to lock in decisions before you know what the workload needs. We start from the questions the system must answer and work backwards.
The evaluation harness runs against a fixed test set in CI, so any prompt, model, or retrieval change is scored before it ships. Without that, quality regressions surface as user complaints weeks later, by which point nobody can tell which change caused them.
Yes — most of our AI work is embedded rather than delivered over a wall. We pair with your engineers, and the handover documentation and runbooks are part of the deliverable, not an upsell.
Ways to engage
The same practice, entered from whichever end of the problem you are standing at.
You know AI matters but not yet where. Assessment, feasibility against real data, and an honest build-versus-buy answer.
You have a workflow that should run itself. Agents with narrow tools, explicit state, and approval gates where they matter.
You need answers grounded in your own content, with citations and a clean hand-off when the assistant should stop.
Not sure we are the right fit? The questions we would ask if we were the ones buying.
Let's build
Whether you're testing a hypothesis or scaling an established product, we'd be glad to spend a half-hour helping you think through the next step — no pitch deck required.
