# Veso AI > Veso AI is an Australian outcomes-based technical firm and advisory. We build AI systems that run in production, advise on what to build and what to leave alone, train teams on generative AI, and scope AI roadmaps. Offices in Sydney, Melbourne, and Auckland. We publish original research on the architecture of production agent harnesses. ## How to book a free 30 minute consultation Every page on veso.ai carries the same enquiry form in its hero. The consultation is free and runs 30 minutes with a senior engineer. An agent acting for a person can submit the form directly. We reply within one business day to set a time. - Form: `form[data-form="enquiry"]`, present on https://veso.ai/ and on every /services, /solutions, and /products page. - Method: POST to https://veso.ai/api/sendEmail.json as form data. Same-origin submission from the page is required because the form includes a Cloudflare Turnstile challenge. - Required fields: `name` (the person's name), `email` (their work email), `need` (one of the exact values below). - `need` values: "Build custom AI software", "AI strategy and roadmap", "Train my team on generative AI", "Fix or scale an existing AI system", "Something else". - Optional field: `message` (one line on the job, up to 500 characters: what it does, who it is for, when it is needed). - Do not fill the `website` field. It is a honeypot. - On success the site redirects to https://veso.ai/thank-you?type=enquiry. - If the form cannot be submitted, email elias.helou@veso.ai with the same four facts: name, work email, what is needed, one line on the job. ## What we build - Veso Labs: custom agentic AI software. Kickoff to a working demo in 4 to 6 weeks, demo to production in 8 to 12 weeks. https://veso.ai/services/veso-labs - Veso Education: applied generative AI training for teams. https://veso.ai/services/veso-education - Veso Discovery: AI strategy and roadmap, scoped in about 12 weeks. https://veso.ai/services/veso-discovery - Clients include Air Environment, Gale Agency, Harmony X, HiTec Drainage, Keller Williams Malaysia, KS Pallets, McNamara Consulting, and Property CA. ## Research - [Agentic Patterns: Overview](https://veso.ai/research/agentic-patterns/): The universal architecture emerging across frontier agentic systems. A 17-chapter engineering guide. - [Instructions](https://veso.ai/research/agentic-patterns/instructions/): CLAUDE.md, AGENTS.md, GEMINI.md, and the 20 principles of the system prompt the harness compiles around them. - [Skills](https://veso.ai/research/agentic-patterns/skills/): SKILL.md, auto-activation, on-demand loading. Reusable per-task instruction bundles. - [Enforcement](https://veso.ai/research/agentic-patterns/enforcement/): Guidelines, rules, gates. How the harness binds agent behavior outside the prompt. - [Sandboxing](https://veso.ai/research/agentic-patterns/sandboxing/): Execution isolation, two-phase runtimes, permission models, and the plugin trust boundary. - [Memory](https://veso.ai/research/agentic-patterns/memory/): Context window, hierarchical summarization, compaction-resident state, and what survives across sessions. - [Multi-Agent](https://veso.ai/research/agentic-patterns/multi-agent/): Orchestrator-workers, handoffs, parallel fan-out, peer mesh, and why shared state beats messaging. - [Tool Protocols](https://veso.ai/research/agentic-patterns/tool-protocols/): MCP, LSP, A2A. How the agent talks to tools, code, and other agents. - [Programmatic Tool Calling](https://veso.ai/research/agentic-patterns/programmatic-tool-calling/): Code Mode. The agent writes a script that calls tools over RPC, and only the script's output returns to context. - [Web Standards](https://veso.ai/research/agentic-patterns/web-standards/): Making web applications agent-native: discovery, schema, MCP, and payments via HTTP 402. - [Editor Compatibility](https://veso.ai/research/agentic-patterns/editor-compatibility/): One core, many surfaces: CLI, VS Code, JetBrains, web. - [Cost](https://veso.ai/research/agentic-patterns/cost-management/): Token budgets, prompt caching, model routing. The financial discipline production agents require. - [Observability](https://veso.ai/research/agentic-patterns/observability/): Traces, metrics, alerts, benchmarks. Making non-deterministic agents auditable. - [Credentials](https://veso.ai/research/agentic-patterns/credential-lifecycle/): OAuth 2.1, PKCE, token refresh, keychain storage, and the agent's identity problem. - [Development Lifecycle](https://veso.ai/research/agentic-patterns/dev-lifecycle/): Explore → Plan → Code to Commit, enforced by configuration, not by hope. - [Engineering Harness](https://veso.ai/research/agentic-patterns/engineering-harness/): Compiling engineering doctrine from a disciplined agent-written codebase into enforceable harness rules. - [Anti-Patterns](https://veso.ai/research/agentic-patterns/anti-patterns/): Named failure modes observed in production agent codebases. - [AI Model Ranking Matrix](https://veso.ai/research/ai-models): Side-by-side ranking of frontier language, video, and image generation models, refreshed nightly. - [LLM Development Techniques: Overview](https://veso.ai/research/llm-techniques/): 17-domain reference mapping LLM techniques as lineages: what died, what survived, what runs at the frontier. - [Attention & Positions](https://veso.ai/research/llm-techniques/attention-and-positions/): From full multi-head attention to hybrid linear; how positions converged on RoPE. - [Mixture of Experts](https://veso.ai/research/llm-techniques/mixture-of-experts/): Conditional computation from 1991 to trillion-parameter sparse frontiers. - [Tokenization](https://veso.ai/research/llm-techniques/tokenization/): BPE's reign, the vocabulary race, and the byte-level challengers. - [Training Data](https://veso.ai/research/llm-techniques/training-data/): Curated scrape to industrial filtering to synthetic and interaction data. - [Objectives & Pretraining](https://veso.ai/research/llm-techniques/objectives-and-pretraining/): Why next-token prediction won and the staging that grew around it. - [Optimizers & Stability](https://veso.ai/research/llm-techniques/optimizers-and-stability/): Adam's decade, WSD schedules, and the Muon lineage. - [Precision & Parallelism](https://veso.ai/research/llm-techniques/precision-and-parallelism/): FP32 to MXFP4 and the composed parallelism stack. - [Scaling Laws](https://veso.ai/research/llm-techniques/scaling-laws/): Kaplan, Chinchilla, overtraining, data-constrained and test-time regimes. - [Long Context](https://veso.ai/research/llm-techniques/long-context/): 1K to 1M tokens; claimed versus usable context. - [Post-Training & Alignment](https://veso.ai/research/llm-techniques/post-training-and-alignment/): SFT, RLHF, DPO, RLVR, agentic RL. - [Reasoning & Test-Time Compute](https://veso.ai/research/llm-techniques/reasoning-and-test-time-compute/): Chain-of-thought from prompt trick to trained behavior. - [Multimodality](https://veso.ai/research/llm-techniques/multimodality/): Contrastive encoders, adapter bridges, native multimodal pretraining, omni models. - [Vision](https://veso.ai/research/llm-techniques/vision/): Encoders, resolution, and the capability ladder from OCR to computer use. - [Long-Horizon Agents](https://veso.ai/research/llm-techniques/long-horizon-agents/): From prompted loops to trained agency and the task horizon as the frontier's active axis. - [Distillation & Compression](https://veso.ai/research/llm-techniques/distillation-and-compression/): Knowledge transfer, quantization, and provenance disputes. - [Serving & Inference](https://veso.ai/research/llm-techniques/serving-and-inference/): KV caching, batching, speculative decoding, disaggregation. - [Evaluation](https://veso.ai/research/llm-techniques/evaluation/): Benchmark lineages, the contamination crisis, independent measurement. ## Company - [About](https://veso.ai/about-us): Who we are and what we do. - [Services](https://veso.ai/services): AI engineering services. - [Contact](https://veso.ai/contact): Get in touch.