Curiosity as proof of capability.

Not everything we build starts as a client project. We keep ideas running in the background—prototypes, technical experiments, benchmarks, and things we simply want to understand.

// r&d_philosophy

Why Scarif Labs maintains an active R&D lab

The software industry is crowded with agencies that wait for a client brief before learning how a new technology works. We believe the best product studio is one that prototypes continuously. We explore edge computing, local-first data convergence, and autonomous agent harnesses before our clients need them, so when production stakes are high, our decisions are grounded in empirical reality.

exp-01Distributed Systems & Infrastructure
7 min read

Edge-First Architectures: Latency as a Foundational Constraint

“What if every computation happened at the edge by default?”

Traditional cloud architectures force every user interaction through centralized data centers located thousands of miles away. By moving execution, caching, and state validation to distributed edge runtimes, latency ceases to be an afterthought and becomes a primary design constraint that transforms user responsiveness.

Core finding:

↳ Speed of light is an immutable physical constraint: moving code closer to the user is the only way to eliminate geographic latency.

Published: 2025-01-15Read full analysis→
exp-02Distributed Systems & Collaboration
9 min read

Local-First Collaboration: Decentralized Consistency with CRDTs

“Can real-time sync work without a central server?”

Local-first software treats the user's local device as the primary source of truth rather than a dumb terminal waiting on remote cloud approval. By pairing local storage with Conflict-free Replicated Data Types (CRDTs), applications achieve zero-latency local interactions and guaranteed mathematical convergence.

Core finding:

↳ Local-first software provides instantaneous zero-millisecond UI latency because every operation modifies local memory first.

Published: 2025-02-01Read full analysis→
exp-03Human-Computer Interaction & AI UX
6 min read

Ambient Computing Interfaces: Background Telemetry and Proactive Systems

“What happens when interfaces respond to presence, not clicks?”

Modern software overwhelms users with thousands of buttons, forms, and notification banners. Ambient computing explores an alternative paradigm: software that observes background context, infers user intent, and prepares answers or automations before the user is forced to ask.

Core finding:

↳ The highest-leverage software interface is often the one that never requires a click.

Published: 2025-02-18Read full analysis→
exp-04Artificial Intelligence & Agentic Systems
8 min read

Production AI Agents: Deterministic Tooling, Evals, and Structured Workflows

“How do we make LLM agents reliable enough for enterprise production?”

Most AI agent demos collapse the moment they encounter real-world ambiguity, rate limits, or unexpected outputs. Engineering production-ready agents requires treating LLMs as probabilistic calculation units inside a deterministic, strongly typed software harness.

Core finding:

↳ Treat language models as probabilistic components wrapped inside deterministic type-safe harnesses.

Published: 2025-02-28Read full analysis→
exp-05Empirical AI Benchmarking · Decision Systems
11 min read · Empirical Benchmark Report

Probabilistic Decision Primitives Under Distribution Shift: An Empirical Benchmark of JEV

“Can a general-purpose probabilistic decision primitive safely extract and retain operational decision signal from software-change metadata under distribution shift?”

We evaluated whether a general-purpose probabilistic decision primitive can extract and retain useful decision signal from software-change metadata under distribution shift. Using dependency updates as the concrete test case, JEV (TypeSafe's System One decision model) was compared against a deterministic static policy and a strong general LLM (DeepSeek Flash) on identical pre-merge state. On a 1,102-case in-distribution benchmark (551 breaking / 551 control), JEV showed substantially stronger ranking quality (AUROC 0.851) than static rules (0.602) and DeepSeek Flash (0.585). On an independently constructed 185-case out-of-distribution (OOD) set (100 breaking / 85 control, 103 independent repositories), JEV retained the highest AUROC among the three (0.605, 95% CI 0.525–0.685), but absolute performance fell sharply. The critical operational finding is negative: JEV's frozen in-distribution threshold did not transfer safely—the 0.62 threshold yielded 30 auto-merges at 50.0% precision with 15 unsafe merges. These empirical results demonstrate that while structured decision models can capture strong ranking signal, probability calibration and threshold stability remain unresolved under domain shift.

Core finding:

↳ Ranking signal does not equal operational safety: high AUROC (0.851) does not prevent unsafe actions if calibration decays under distribution shift.

Published: 2026-09-25Read full analysis→
Start a project

Have a hard engineering challenge?

When you bring us an ambitious problem, we've already explored the terrain.

Start a project hi@scariflabs.com