AI development services: agents that look up company data and carry out tasks in your systems, with tested quality and usage cost under control.
Free first conversation · Reply within 1 business day
DIAL, for The Hackett Group (NASDAQ: HCKT):
74pages and 27 API routes on Claude
Mirador:
Generative AI with Claude and OpenAI behind a provider abstraction
Zorbit (own product):
AI coding agents orchestrated with ~74 MCP tools
Contas Casal (own product):
LLM assistant with RAG and a semantic cache
DIAL, for The Hackett Group (NASDAQ: HCKT): 74 pages and 27 API routes on Claude
Mirador: generative AI with Claude and OpenAI behind a provider abstraction
Zorbit (own product): AI coding agents orchestrated with ~74 MCP tools
Contas Casal (own product): LLM assistant with RAG and a semantic cache
The agent impresses in the demo but fails on real data, and nobody can measure answer quality.
The agent gets lost in long tasks: it forgets what it already decided, fills the context window and repeats expensive model calls.
The LLM bill grows every month, with no visibility of cost per flow, model or customer.
Security and legal are holding the launch over prompt-injection and data-leak risk.
Agents that plan long tasks, act on your systems (tool calling), delegate parts to subagents with isolated context and reuse skills, with human approval at decision points.
We design the harness around the model (loop, tools, permissions, rules and checks) and what enters the context: short-, medium- and long-term memory, compaction, token budget and handoffs.
A single agent or a multi-agent system, with routing between agents: we choose LangChain, LangGraph, Deep Agents, Mastra or the Vercel AI SDK for the problem and your team's stack.
The model answers from company documents and databases and cites the source: chunking, embeddings, a vector store (pgvector), hybrid search, reranking and structured queries for tabular data.
MCP (Model Context Protocol) is an open standard for connecting agents to systems: servers that expose ERP, CRM and APIs as tools and resources, with per-tool permissions and audit trails.
Model swaps by configuration (Claude, OpenAI, Gemini), routing per task and fallback across providers; prompt caching and semantic caching cut latency and token cost.
A versioned dataset of real and edge cases, evals in CI with LLM-as-a-judge, and traces in OpenTelemetry, LangSmith or Langfuse: quality, latency and cost per task on every change.
Input and output validation, sanitized LLM responses and adversarial tests before every release, so the agent neither leaks data nor follows instructions hidden in the content it reads.
Supervised models for classification, regression and forecasting, validated on held-out data; fine-tuning and reinforcement learning (RLHF, RL with verifiable rewards) only when an eval shows a gain.
See also: AI for e-commerce, software architecture consulting, and AI training for engineering teams.
DIAL: architecture and delivery of the AI-native data-strategy platform, on Anthropic Claude.
Mirador: generative-AI application engine, with Claude and OpenAI behind a provider abstraction.
Facing a similar challenge? Tell us the problem and the timeline.
Free first conversation · Reply within 1 business day
AI & LLM engineering is AI agent development for companies: assistants built on LLMs, the technology behind ChatGPT, that look up company data and carry out tasks in your systems, with quality tests, safety limits and cost under control. Hire it when the AI use case is already clear and what is missing is making it run safely.
We read the code, the data and the business goal, and map use cases, integrations and risks. You receive an assessment with the target architecture, the plan and the estimate.
Specs, ADRs and acceptance criteria before the first line of code. We build the evaluation dataset: real questions, expected answers and adversarial cases. You approve what will be measured.
One end-to-end flow on real data, measured by the evals. You receive a working PoC, the evaluation results and a recommendation to proceed, adjust or stop.
Short iterations with frequent releases: integrations, guardrails, evals, traces and minimum test coverage as a condition for every release. You receive the system in production, with CI/CD.
We track quality, latency and cost through the traces, document the system and prepare your team to run it. You receive the documentation and the dashboards; evolution can continue under a monthly retainer.
| Scope | Typical timeline | Engagement model |
|---|---|---|
| AI discoveryUse cases, data, target architecture, plan and estimate | 1–2 weeks | Fixed-scope project |
| Proof of conceptOne end-to-end flow on real data, with a dataset and evals | 2–4 weeks | Fixed-scope project |
| ProductionAgent, RAG or MCP server integrated with your systems, with guardrails, evals, traces and CI/CD | 6–12 weeks | Fixed scope or time & materials with a cap |
| Ongoing evolutionNew flows, with quality and cost overseen by a part-time (fractional) AI architect | Ongoing | Monthly fractional retainer |
Timelines are typical ranges. Discovery defines the scope and the estimate of each project; the contract can be in USD or BRL.
Want to know which format fits your project?
Free first conversation · Reply within 1 business day
We build AI products and we use AI to build them. The same rules apply to both: specs before code, a harness with rules and checks, automated evals and human review of decisions.
Where AI comes inAI coding agents (Claude Code, Codex, Cursor, Copilot, OpenCode) implement from approved specs, following Spec-Driven Development, inside a harness of rules, skills and quality hooks.
What humans reviewArchitecture, ADRs, acceptance criteria and every merge go through senior human review; in the product, sensitive agent actions require human approval.
Quality gatesA merge only happens with automated tests passing, the coverage target met and CodeQL static analysis passed.
Evals in CIThe evaluation dataset runs on every prompt, model or tool change, with code checks and LLM-as-a-judge; a quality regression blocks the release.
Traces in productionEvery run becomes a trace (OpenTelemetry, LangSmith or Langfuse) with the model calls, tools, tokens and cost of each step.
Guardrails and dataValidated and sanitized LLM output, prompt-injection defenses and per-tool limits; client data never trains models, and we work under NDA.
| Approach | When to use it | What it requires | Main risk |
|---|---|---|---|
| RAG | The model must answer from company data that changes often | Chunking, embeddings, hybrid search, reranking and evals of retrieval and answers | Poor retrieval yields a wrong answer that looks right |
| Fine-tuning | Stable tone, format or task, when an eval shows a gain over prompt and context | A curated dataset, evals before and after, and retraining on every change | Stale knowledge and maintenance cost |
| Tool-using agent (MCP) | The model must act on systems: query, create or update records | A harness with scoped permissions, autonomy limits and auditing | Unintended actions without guardrails or human approval |
| Fixed workflow with an LLM | The steps are known and the model only classifies, extracts or summarizes | Deterministic orchestration and tests per step | Little flexibility outside the planned path |
| Supervised machine learning | Classify, forecast or estimate from labeled historical data | Labeled data, data science and validation on held-out data | Accuracy drops when the data pattern shifts |
In most cases, RAG: it queries current company data on every question, cites the source and needs no retraining when information changes. Fine-tuning pays off when the problem is tone, format or a very specific, stable task, and only when an eval shows a gain over prompt and context. Combining both is common, and discovery shows which approach your case calls for.
Context engineering is deciding what enters the model's window at each step: short-, medium- and long-term memory, compaction, the token budget and handoffs between agents and sessions, so the agent is more accurate and costs less. Harness engineering is everything around the model, such as tools, permissions and checks; the term has its own guide: “What Is Harness Engineering for AI Agents?”.
It depends on the problem and on your team's stack, and the choice is recorded in an ADR. LangGraph fits stateful flows with persistence and human approval, designed as a graph; Deep Agents, built on LangGraph, adds planning, subagents and skills for long tasks; LangChain provides the building blocks; Mastra and the Vercel AI SDK suit TypeScript teams. For autonomous AI agents such as OpenClaw and Hermes Agent (Nous Research), we set up permissions, channels and scheduled tasks.
With a versioned dataset of real and edge cases, automated evals on every prompt, model or tool change, and a trace of every run in production. Evals combine code checks, human review and LLM-as-a-judge, and a regression blocks the release. Traces in OpenTelemetry, LangSmith or Langfuse show accuracy, latency, tokens, cost per task and hand-offs to a person. For an agent that already exists, this measurement is the first step of discovery.
Yes: when the problem needs its own model, we apply data science and supervised learning (classification, regression and forecasting), validated on held-out data. LLM fine-tuning and reinforcement learning (RLHF, RL with verifiable rewards) come in only when an eval shows a gain over prompt and context, because they cost more to maintain. We do not train foundation models: our focus is solving business problems.
With layered defenses: system instructions isolated from external content, input and output validation, sanitized LLM responses and least-privilege permissions per tool. Sensitive actions require human approval, and adversarial tests run in CI before every release. On Mirador, The Hackett Group's generative-AI engine, we implemented prompt-injection defenses and an audit trail of the system prompt.
It depends on three factors: the number of integrations, the quality of the data and how much autonomy the agent has. That is why we do not publish a price list. A 1–2 week discovery defines scope, risks and estimate, and the proposal can be fixed scope or time & materials with a cap, in USD or BRL. Beyond development there is the recurring token cost, which we forecast, monitor per model and reduce with prompt caching and semantic caching.
Master Timm builds production AI agents, RAG systems and MCP servers from São Paulo, Brazil, for companies in the United States and Brazil. For The Hackett Group (NASDAQ: HCKT), we designed and delivered the DIAL data platform and Mirador, a generative-AI engine on Claude and OpenAI; both are in our case studies. Whoever designs the architecture also writes the code, with nothing handed off to a junior team.
AI-focused IT consultancy
Senior engineering from discovery to delivery: whoever designs the architecture also writes and reviews the code.
Master Timm · Updated on
In the first conversation we look at the use case, the data and the integrations, and tell you plainly whether it calls for an agent, RAG, a machine learning model or none of them.
Free first conversation · Reply within 1 business day