Advisory
For teams building production agents.
I help teams reason through context architecture, agent reliability, evaluation, prompt engineering, and the translation layer between domain experts and engineering. The useful work is usually before scale: naming the actual failure modes, deciding what context belongs where, and turning expert judgment into something the system can use.
Email muratcan.koylan@outlook.com with a short note: what you are building, who uses it, what currently breaks, and what decision you need to make.
Good Fit
I am useful when
- You are building production agents, but quality degrades when real users, messy context, or handoffs enter the system.
- Your team has prompts, tools, and documents scattered across the stack, and no clear context architecture.
- You need evaluation rubrics that reflect expert judgment instead of generic thumbs-up / thumbs-down QA.
- Your non-technical team knows the workflow deeply, but engineering is missing the tacit rules that make the work good.
- You are deciding whether to build, buy, fine-tune, or wrap an AI tool around an existing business process.
Tracks
Two different advisory surfaces
For AI engineering and product teams
Context architecture review
Audit what enters the model: instructions, files, memory, tools, retrieved context, user state, and intermediate agent outputs.
Multi-agent system design
Review when to use handoffs, when to use tool-style consultation, and where a single agent with better tools is the simpler answer.
Evaluation and rubrics
Turn expert judgment into scoring frameworks, acceptance criteria, test cases, and feedback loops that can be inspected.
Prompt engineering systems
Review instruction hierarchy, dynamic prompt composition, tool protocols, and prompt patterns that need to hold under production use.
Memory and routing systems
Design what should persist, where it should live, how long it should survive, and what should be ignored.
For marketing and ops teams adopting AI
AI vendor evaluation
Separate useful capability from polished walkthroughs. Define tests around your real data, workflow constraints, and failure tolerance.
Marketing and ops workflow design
Map where AI should assist, where humans should stay in the loop, and which manual steps are worth automating.
Expert knowledge extraction
Interview operators, strategists, or domain experts and translate their implicit rules into usable agent instructions and rubrics.
GTM and research automation
Design research, enrichment, outbound, and reporting workflows that use agents without turning the stack into fragile glue code.
Engineering-to-operator translation
Help technical and non-technical teams define the same problem, constraints, and success criteria before implementation starts.
Engagement Shapes
Ways to work
System review
A focused review of prompts, agent flow, context loading, tools, evaluation, and failure cases. Usually starts with docs, traces, and one live walkthrough.
Advisory sprint
A short working block around one problem: agent reliability, persona design, evaluation, memory routing, vendor assessment, or workflow architecture.
Workshop
A practical session for engineering, marketing, ops, or leadership teams. Best when the team brings a real workflow or agent to dissect.
Ongoing advisory
Periodic review for teams already building. Useful when architecture, product decisions, and operator feedback are moving at the same time.
Grounding
What this is based on
Clinical AI context systems
At Sully.ai, I work on context engineering for AI Scribe and agent harnesses across medical specialties.
Marketing expert agents
At 99Ravens AI, I built multi-agent architecture, persona layers, agent skills, prompt engineering systems, and evaluation rubrics for strategy workflows.
Agent Skills for Context Engineering
Open-source patterns for context, memory, tools, evaluation, and multi-agent systems. Cited in Meta Context Engineering and mapped in the Agent Harness Engineering survey.
GTM and research systems
Built AI SDR, research, and enrichment workflows across Clay, Apollo, HubSpot, Firecrawl, Make, and LLM pipelines.
Specialized model work
Fine-tuned an 8B memory-routing model and built reusable pipelines for style transfer and task-specific model training.
Not A Fit
Where I am probably not useful
- Generic prompt packs with no workflow, data, or evaluation plan.
- Growth hacks where the goal is volume rather than system quality.
- Replacing your engineering team. I can review architecture and help shape the system, but implementation still needs owners.
- Compliance certification, legal review, or security audit work that requires a formal vendor.