Denis Baciu https://denisbaciu.com/ Writing on industrialising AI in financial services — taking proof-of-concepts into production. en Fri, 04 Sep 2026 16:13:21 +0000 Why I'd instrument before building an internal AI cluster https://denisbaciu.com/writing/2026-08-i-d-instrument-before-building-internal.html https://denisbaciu.com/writing/2026-08-i-d-instrument-before-building-internal.html Sun, 30 Aug 2026 00:00:00 +0000 A practical look at Nutanix's $20M AI cluster and the build vs buy decision for heavy SaaS AI spend. The coordination tax in a polyrepo estate https://denisbaciu.com/writing/2026-08-coordination-tax-polyrepo-estate.html https://denisbaciu.com/writing/2026-08-coordination-tax-polyrepo-estate.html Sat, 29 Aug 2026 00:00:00 +0000 A polyrepo estate behaves like a distributed system, and recurring coordination work should drive the monorepo decision, not ideology. Open-weight models are moving into production workloads https://denisbaciu.com/writing/2026-08-open-weight-models-moving-production-workloads.html https://denisbaciu.com/writing/2026-08-open-weight-models-moving-production-workloads.html Thu, 27 Aug 2026 00:00:00 +0000 A practical look at evaluating open-weight models on real traffic, piloting them in low-risk tasks, and choosing managed inference. Slack Code moves AI coding out of the terminal into group chat https://denisbaciu.com/writing/2026-08-slack-code-moves-ai-coding-out.html https://denisbaciu.com/writing/2026-08-slack-code-moves-ai-coding-out.html Mon, 24 Aug 2026 00:00:00 +0000 Slack Code puts coding agents in shared channels so teams can prompt, review, and steer AI-generated changes with a searchable record. Boring Infrastructure Sleeps Better https://denisbaciu.com/writing/2026-08-boring-infrastructure-sleeps-better.html https://denisbaciu.com/writing/2026-08-boring-infrastructure-sleeps-better.html Sun, 23 Aug 2026 00:00:00 +0000 Rogue Amoeba's move from CGI to FastCGI and MySQL to SQLite shows how deliberately unglamorous upgrades can reduce operational pain and improve resilience. Enterprise AI’s Quiet Shift from Experiments to Unit Economics https://denisbaciu.com/writing/2026-08-enterprise-ais-quiet-shift-experiments-unit.html https://denisbaciu.com/writing/2026-08-enterprise-ais-quiet-shift-experiments-unit.html Sat, 22 Aug 2026 00:00:00 +0000 Enterprise AI adoption is moving from broad experimentation to disciplined, measurable use cases driven by cost observability and per-query ROI. An LLM wiki is not a self-maintaining knowledge base https://denisbaciu.com/writing/2026-08-llm-wiki-not-self-maintaining-knowledge.html https://denisbaciu.com/writing/2026-08-llm-wiki-not-self-maintaining-knowledge.html Thu, 20 Aug 2026 00:00:00 +0000 Casey Newton's LLM wiki shows real research gains but also ongoing maintenance, where curation is the real product. MCP session handles are capability tokens, not pointers https://denisbaciu.com/writing/2026-08-mcp-session-handles-capability-tokens-not.html https://denisbaciu.com/writing/2026-08-mcp-session-handles-capability-tokens-not.html Mon, 17 Aug 2026 00:00:00 +0000 MCP's shift to model-held session handles trades server affinity for horizontal scale; short-lived capability tokens keep both. Why Regulated Enterprises Will Mix Open and Closed Models https://denisbaciu.com/writing/2026-08-regulated-enterprises-will-mix-open-closed.html https://denisbaciu.com/writing/2026-08-regulated-enterprises-will-mix-open-closed.html Sun, 16 Aug 2026 00:00:00 +0000 Regulated enterprises will increasingly blend open-weight and closed models for AI agents, driven by control, auditability, and customization rather than benchmark scores. Multi-cloud AI observability needs a common transponder https://denisbaciu.com/writing/2026-08-multi-cloud-ai-observability-needs-common.html https://denisbaciu.com/writing/2026-08-multi-cloud-ai-observability-needs-common.html Sat, 15 Aug 2026 00:00:00 +0000 How OpenTelemetry and AWS AgentCore Observability can turn fragmented AI agent monitoring into one operational pane across clouds. AI-assisted PostgreSQL migrations need a human review loop https://denisbaciu.com/writing/2026-08-ai-assisted-postgresql-migrations-need-human.html https://denisbaciu.com/writing/2026-08-ai-assisted-postgresql-migrations-need-human.html Thu, 13 Aug 2026 00:00:00 +0000 Google Cloud's Gemini in Database Migration Service drafts PL/pgSQL conversions, but teams must keep a strict review loop to avoid subtle logic changes. The Real Bottleneck in Data Platforms Is Neglected Orchestration https://denisbaciu.com/writing/2026-08-real-bottleneck-data-platforms-neglected-orchestration.html https://denisbaciu.com/writing/2026-08-real-bottleneck-data-platforms-neglected-orchestration.html Mon, 10 Aug 2026 00:00:00 +0000 While tool sprawl gets the blame, the true problem is treating orchestration as a cron job instead of a control plane with SLOs for freshness, failure recovery, and cost. Why I’d Bet on the Agent Access Model for Production https://denisbaciu.com/writing/2026-08-id-bet-agent-access-model-production.html https://denisbaciu.com/writing/2026-08-id-bet-agent-access-model-production.html Sun, 09 Aug 2026 00:00:00 +0000 Cloudflare’s Agent Access Model authorizes each agent action at the task level, using short-lived credentials and a Trust Ratchet to eliminate privilege creep. Memory is the bottleneck nobody budgets for https://denisbaciu.com/writing/2026-08-memory-bottleneck-nobody-budgets.html https://denisbaciu.com/writing/2026-08-memory-bottleneck-nobody-budgets.html Sat, 08 Aug 2026 00:00:00 +0000 CXL memory pooling relieves GPU HBM constraints in RAG and vector search by offloading KV caches and embeddings to shared low-latency memory. The enterprise AI strategy that outlasts any single model https://denisbaciu.com/writing/2026-08-enterprise-ai-strategy-outlasts-any-single.html https://denisbaciu.com/writing/2026-08-enterprise-ai-strategy-outlasts-any-single.html Thu, 06 Aug 2026 00:00:00 +0000 An effective AI strategy abstracts providers, evaluates on real tasks, monitors costs and latency, and avoids vendor lock-in. Agents Are Easy to Build, Hard to Control https://denisbaciu.com/writing/2026-08-agents-easy-build-hard-control.html https://denisbaciu.com/writing/2026-08-agents-easy-build-hard-control.html Mon, 03 Aug 2026 00:00:00 +0000 Real-time observability and governance must be built into AI agents from day one, or organizations face black-box risks and spiraling costs. Overprivileged AI Agents Repeat Old IAM Mistakes https://denisbaciu.com/writing/2026-08-overprivileged-ai-agents-repeat-old-iam.html https://denisbaciu.com/writing/2026-08-overprivileged-ai-agents-repeat-old-iam.html Sun, 02 Aug 2026 00:00:00 +0000 71% of organizations have AI agents that can reach sensitive data; one-third of developers saw incidents from overprivileged non-human identities. Stateful Execution Is the Next Bottleneck for Enterprise AI Agents https://denisbaciu.com/writing/2026-08-stateful-execution-next-bottleneck-enterprise-ai.html https://denisbaciu.com/writing/2026-08-stateful-execution-next-bottleneck-enterprise-ai.html Sat, 01 Aug 2026 00:00:00 +0000 Stateful execution environments are set to become the primary bottleneck for enterprise AI agents, with Perplexity’s SPACE as an early signal. Observability Is a Safety Net, Not a Silver Bullet, for Vibe Coding https://denisbaciu.com/writing/2026-07-observability-safety-net-not-silver-bullet.html https://denisbaciu.com/writing/2026-07-observability-safety-net-not-silver-bullet.html Thu, 30 Jul 2026 00:00:00 +0000 Observability exposes gaps in AI-generated code, but its feedback loop is slow; enforce contracts and use runtime data for targeted fixes. Routine coding tasks don't need expensive reasoning models https://denisbaciu.com/writing/2026-07-routine-coding-tasks-don-t-need.html https://denisbaciu.com/writing/2026-07-routine-coding-tasks-don-t-need.html Mon, 27 Jul 2026 00:00:00 +0000 Routine coding tasks don't need deep chain-of-thought, and using expensive models for them wastes money; a router sending simple prompts to cheap models solves that.