Bernard Kiyanda

We just published a new post on the Tensor9 blog about a problem that comes up in almost every vendor conversation we have: what happens when your application depends on a managed service like DocumentDB, and your customer can't run on AWS. https://lnkd.in/gzamYh2d

9

1 Comment

"Reliability" used to mean uptime. Eugene Wu points out that for agents, the definition of reliability is broader: did the agent understand the request and its intent, and then do the right thing? He calls it semantic coupling: a small retrieval miss early in the chain cascades quietly into a confident, wrong answer. My conversation with Eugene, associate professor of computer science at Columbia University and co-director of the Columbia Data Agents Process Lab (DAPLab), is live on AI Radicals. Full episode link in the comments. #AgenticAI #AIReliability

7

2 Comments

Excited to share that Sphere XG now includes an updated version of our beloved built-in orthophoto creation and export - directly from your point clouds! Thanks to your feedback, the new solution now produces faster, cleaner, and more map-accurate, high res imagery, all without leaving the platform. A big step forward for our users in mapping and digital twins! #SphereXG #orthophoto #geospatial #innovation

16

 I shipped MoJoAssistant v1.2.1-beta this week. Two things I'm particularly happy with:  The Attention Layer. When an agent running in the background needs a decision — which subnet to scan, which file to modify, what to do next — it now surfaces that question the moment you start a conversation, not buried in a log somewhere. A deterministic classifier assigns every system event a priority level (0–5) at write time. No LLM inference, no latency. The orientation response injects anything blocking or alerting automatically, and stays silent when everything is quiet. MCP tool consolidation: 49 → 12. One of the subtler problems with giving an LLM a large tool list is that management tools (restart daemon, remove document, revoke resource) compete for attention with everyday tools (save conversation, reply to a task). The new design keeps five frequent tools visible at all times and puts everything else inside seven action hubs. Each hub is self-describing — call it with no arguments and it returns its own help menu. Unknown actions self-correct the same way. The LLM never needs to guess.  Also shipped: Human-in-the-Loop inbox with zero-latency wake signal, config-driven scheduler defaults, push notifications (ntfy), role policy enforcement, config doctor, and a  93-case integration smoke test covering every tool — happy paths and expected failures — that runs in CI before every release. MoJoAssistant is a local-first privacy proxy: a trusted intermediary where a local thinking LLM handles sensitive data before anything touches an external service. v1.2.1 is the attention and surface layer. Data boundary enforcement and PII classification are next. GitHub: https://lnkd.in/gZiBKc22

4

We’ve moved beyond simple autocomplete—GenAI, agents, and chat-driven workflows are changing how real software gets built. This book - “Vibe Coding” by Gene Kim and Steve Yegge is about turning AI from cool demos into production-grade engineering. It shows how to work with AI effectively, put the right guardrails in place, and maintain quality, security, and engineering discipline while moving much faster. Perfect holiday read if you’re serious about how to use AI like a pro, not just why. 2026 is coming—time to sharpen the tools.

32

2 Comments

Per-Step Accuracy % vs End to End Success: 10 Step Workflows 90 % Accuracy => 65.1 failures in 100 workflows 99.9% Accuracy => 1 failure in 100 workflows

1