Łukasz Jaźwa

AI coding assistants have quietly become a real line item in engineering budgets — and most teams have no strategy for it, just a bill. The uncomfortable truth: you don't pay for tokens, you pay for tokens that produce no value.

This talk is a practical playbook built on two levers. First, model routing: matching model tiers to task phases — frontier models for planning and review, mid-tier for implementation, the cheapest for mechanical work. Second, context management: why agents that explore large codebases burn premium tokens on searching instead of thinking, and the architectural pattern that fixes it — a code intelligence layer where an expensive model asks questions and a cheap, locally-run system that already knows the codebase answers them.

Main points covered:
- Why AI spend is a value problem, not a price problem — and the two levers that control it: which model does the work and what's in its context
- The Plan–Build–Grind–Review routing framework: frontier models for planning and review, mid-tier for implementation, the cheapest tier for mechanical work
- Where specialized and local models fit: embeddings, retrieval, and privacy-sensitive workloads at near-zero marginal cost
- The false-economy trap: cost × success rate — how retries and failed reviews make the cheapest model the most expensive one
- Context as the bigger lever: agents burning frontier-model tokens on codebase exploration, and why long contexts degrade reasoning
- The code intelligence layer pattern: the expensive model asks questions, a cheap locally-run system that knows the codebase answers them
- The handful of signals that tell you the system is working

Key takeaways:
- A routing framework (Plan–Build–Grind–Review) for matching model tiers to engineering task phases, with the cost metric that actually matters.
- The code intelligence layer pattern: put retrieval in front of your agents so premium models think instead of search — cutting cost and context rot at once.
- A mental model for treating token economics as an architecture concern — deciding what belongs in a model's context, not just which model to call.

Why is this interesting for listeners?
Every engineer in the room is already using AI assistants — and someone on their team is watching the bill grow with no plan behind it. This talk hands them a concrete framework they can apply Monday morning to cut that spend without slowing down, turning a vague budget anxiety into a set of architectural decisions they control. It reframes "AI is expensive" as "we're using it wrong" — a message developers can act on themselves.

Łukasz Jaźwa is the CTO of Synergy Codes, a 70-person software company from Wrocław, Poland, specializing in diagramming and workflow tooling, including ngDiagram, an Angular diagramming library. He leads AI tooling adoption across the entire engineering organization: Claude Code, Cursor, and agentic workflows, plus the unglamorous parts like token monitoring, internal plugin distribution, and getting 70 engineers to actually change how they work.

He is also co-founder of Tentaro, an on-premise code context platform that indexes repositories and serves relevant context to AI models. Tentaro was born from Synergy's own pain: agents burning tokens on repeated exploration of large codebases. Łukasz pays the AI bills he talks about, and he built the system that shrank them.

Outside of work he is into physics, history, and hiking in the Polish mountains.