Farshid M.

I'm the senior AI architect founders and CTOs bring in when an AI project has to actually work in production — and there's no second chance to get it right. Twenty-five years of engineering. Founder of two AI companies (one bootstrapped to 50+ enterprise customers; the other a pre-production AI agent validation platform with active POCs). Three USPTO patents on AI automation. Cited in the Gartner® Hype Cycle™ for Security Operations 2024. Spoke at PyTorch Conference 2025 on agentic AI systems. I currently advise senior executives at publicly traded companies on AI adoption and architecture. Most of my Upwork engagements are short, high-leverage, and decision-shaping rather than long-running implementation. What I'm typically brought in for: → Agentic AI systems that need to leave the demo stage. Multi-agent orchestration (LangGraph, AutoGen, CrewAI, Semantic Kernel, MCP), tool design, planning loops, memory architecture, agent evaluation, and the failure modes that kill production deployments. I've shipped a Rust-native swarm of hundreds of autonomous agents for a real product, so I know where the seams give way. → LLM systems architecture. RAG that holds up under load (hybrid retrieval, eval harnesses, SLOs, rollback paths), fine-tuning strategy (when to fine-tune, when not to, what to measure), prompt and context engineering at the system level, model routing, cost and latency tradeoffs. → AI platform and infrastructure decisions. Whether to build or buy, which inference stack (vLLM, Triton, Bedrock, hosted), how to design GPU capacity, how to architect for SOC2 / FedRAMP / regulated environments, how to set up observability and evals that actually catch regressions. → AI security, governance, and red-teaming. Prompt injection defense, agent security review, audit logging design (including post-quantum-ready cryptographic chains), responsible AI guardrails, and validation frameworks for AI systems that need to satisfy auditors. → Senior technical advisory and CTO-for-hire engagements. Architectural reviews, hire-vs-build calls, technical due diligence, AI strategy for non-AI-native companies, executive coaching on AI adoption. How I work with clients: Most engagements look like one of three shapes. Architectural assessment — I spend two weeks reading your code and talking to your team, and deliver a written assessment of what's load-bearing, what's at risk, and the prioritized sequence of changes. Targeted build — a tightly scoped 4-8 week engagement with defined deliverables and weekly written progress. Senior advisory — recurring hours, typically two to four per week, where your team brings me the hard decisions and I help you make them faster than you would alone. I don't take work that requires 24/7 on-call, work that's better done by a junior at a lower rate, or work where the scope is "build me an AI." I'll tell you in the first call if I'm not the right fit and point you to someone who is. Stack I work in: Python, Rust, Go · PyTorch, Hugging Face (Transformers, TRL, PEFT) · LangGraph, AutoGen, CrewAI, Semantic Kernel, MCP · LangChain, LlamaIndex · Weaviate, Qdrant, pgvector · vLLM, LiteLLM, Axolotl, Ollama · DeepSpeed, Flash Attention, QLoRA · MLflow, Langfuse, Weights & Biases · AWS Bedrock / SageMaker, Azure OpenAI, GCP Vertex AI · Kubernetes, Terraform · SOC2, FIPS 140-2, FedRAMP, NIST. To start a conversation: send three things — what you're building, the specific decision or problem you'd want me involved in, and the engagement shape you have in mind (assessment, build, or advisory). I'll respond with whether I can help, how I'd approach it, and a proposed first step. If we're not a fit, I'll say so. — Farshid