We're an AI engineering studio.

Use downtime to "think ahead". Precompute, cache, and cut test-time cost 5x while boosting accuracy. Done for BattlePal, our first foray into "proactive" AI.

Multi-turn chats break models - 39% performance drop from context drift and premature answers. We design resets, not endless chats.

Don't dump every tool into the prompt. Retrieve only what matters. Tool selection accuracy x3. We later improved on this using findings from https://manus.im/blog/Context-Engineering-for-AI-Agents-Lessons-from-Building-Manus

Counter-intuitive but true: fewer tools = better outcomes. It made us better in building MCP servers for our agents. Tools shouldn't simply wrap each API endpoint, but correspond to a distinct "workflow

Helped us with on-device deployment of one of our earlier GenAI product, Imms.

Retrieval that explains itself. We now use reasoning traces for debugging and monitoring.

We let SOTA models build tools and use smaller models in production deployments to decrease costs manifold.

For EU GDPR, CCPA compliance we now use data minimization and strong access controls instead of "unlearning". Sensitive data is retrieved, never goes into the training dataset.

Allowed us to experiment with retrieval granularity. "Propositions" oftentimes improved RAG accuracy.

We're surprised this paper is not more widely cited. Thanks to it we optimize our agentic loops for accuracy & cost, simultaneously.

Informed many of our product decisions. Instead of providing ready answers or RAG based solutions we, whenever it makes sense, build systems that help users uncover unknown unknowns or implicit assumptions about tasks.

Enlightning take on knowledge graphs. Contains implementation of Personalized PageRank. We've eventually discarded it in favor of more matured MemGPT and Mem0.

GenAI can bring evaluation overhead, broken workflows and harder hard tasks. We fix it with feedback loops and, emphasized in the paper, Ecological Interface Design (EID).

Deepened our understanding of 'cascade of models'. End-to-end accuracy +70%.

We reference CaMeL (Capability-based Model Layer) in our red teaming efforts.

We can recommend this reading enough. It saves us tons of time on agentic loop architecture design.

This and other surveys is a boundless inspiration for new agent designs.

Despite dramatical title, it expands on MCP-RAG and other techniques to not overwhelm agents with too many, too granular tools modeled after CRUD APIs.