DEV BY SALDEV BY SAL LTD. is a senior software studio building AI-native products, agentic systems and revenue-shifting platforms for ambitious teams. Weeks, not quarters.
Selected partners
A small sample of what shipping looks like with us. Real customers, real metrics, real production traffic.
LLM-backed products with eval harnesses, retrieval, agents, and the boring infrastructure that keeps them honest in production.
Internal platforms that compress weeks of work into hours. Workflow builders, SDKs, CLIs, and the kind of tooling engineers tell their friends about.
A weeks-not-months engagement to ship the thing your roadmap has been failing to ship. Senior engineers, no project managers, real software.
Every engineer on your project has shipped a startup or built infrastructure at a name-brand company. No juniors hidden behind a Slack handle.
Our default engagement is four weeks. If we cannot describe a credible shape in two, we tell you instead of pretending.
We work in your repo, your stack, your conventions. When we leave, your team owns it — not us.
We say no when we should. The shops that say yes to everything are the shops that ship nothing of consequence.
Longer-form pieces for founders, CTOs and engineering leaders — the kind of writing we read when we want to think more clearly about ambitious software.
Strategy papers, evaluation playbooks and engineering notes from production AI work.
Most production LLM systems are evaluated by other LLMs, against benchmarks that have probably leaked into the model's training data, with judges that systematically prefer their own kin. The recent literature is clear-eyed about all three problems. This paper is a working playbook for shipping evaluation infrastructure that survives those facts.
The literature on production LLM agents has matured from "can we make this work at all?" to "which interface and memory structure makes it work reliably?" This paper walks the dominant patterns — ReAct, Reflexion, AutoGen-style multi-agent, and SWE-agent's interface-as-product — and the failure modes that come with them.
Production RAG in 2026 looks nothing like the "embed, retrieve, generate" of 2023. The recent literature converges on retrieval as a routing problem — when to retrieve, how to verify retrieval quality, when to switch to graph indexing, when to lean on long context instead. This paper is the practitioner-facing version.
Shorter pieces from inside the studio. Engineering decisions, post-mortems, and the occasional opinion.
All postsPeople keep asking why we ship in four weeks instead of six months. It's not because we're faster. It's because four weeks is what's actually required to learn anything.
LangChain, LlamaIndex, Haystack — we have used them all and we ship most of our agents without them. Here is when frameworks help, when they hurt, and why we usually choose the harder-looking option.
For LLM-shaped products, the document that defines what the product does is the eval set, not the PRD. We have stopped writing PRDs for AI features and started writing the cases. The work is better for it.