@liarjo

With the mind in the cloud 😎 Opinions are my own. RTs ≠ endorsement

 Boston, MA
jpgarcia.cl
Joined June 2008
  • Tweets6.4K
  • Following408
  • Followers282
1.4KMedia
MIT announces the MIT for America initiative. The effort aims to help U.S. learners from kindergarten to community college, with an emphasis on math, making, and the constructive use of AI. news.mit.edu/2026/mit-americ…

MIT announces the MIT for America initiative, to strengthen STEM education across the country

MIT for America is a new initiative that seeks to strengthen STEM education nationally, at many levels of learning. The Institute-wide effort ranges across STEM fields and will engage students from...

news.mit.edu
38
160
16
767
48,641
Bedrock managed entitlements: subscribe to private pricing once centrally, then auto-distribute and activate License Manager grants across your whole org. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot at a central LICENSES vault distributing keys along branching lines to multiple AWS account boxes, banner reading Managed Entitlements, labels Bedrock and License Manager

ALT Pencil sketch of a retro AI robot at a central LICENSES vault distributing keys along branching lines to multiple AWS account boxes, banner reading Managed Entitlements, labels Bedrock and License Manager

14
Template-based document processing breaks when layouts change. This agentic IDP system on Amazon Bedrock AgentCore uses 8 agents that classify, extract, validate, and learn from documents — new formats go straight to extraction without retraining. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot filing invoices into a Processed Documents cabinet with a Learn-Extract-Adapt-Improve loop, banner reading Self-Improving IDP, labels AgentCore and Learn & Extract

ALT Pencil sketch of a retro AI robot filing invoices into a Processed Documents cabinet with a Learn-Extract-Adapt-Improve loop, banner reading Self-Improving IDP, labels AgentCore and Learn & Extract

12
Amazon CloudWatch Omni: AI-powered observability for the AI era. Eval-driven tracing that answers "why did my agent do that?"—correlating agents, apps & infra in one console. aws.amazon.com/blogs/mt/intr…
Pencil sketch of a retro AI robot at a mission-control desk watching observability dashboards, banner reads CloudWatch Omni

ALT Pencil sketch of a retro AI robot at a mission-control desk watching observability dashboards, banner reads CloudWatch Omni

24
Not every agent decision needs an LLM. Strands Decider is a 1.9B "system one" model that picks options or scores on a scale in ~115 ms, with a calibrated confidence on every answer. github.com/strands-labs/stra…
Pencil sketch of a retro AI robot at a fork in the road choosing between option cards with a confidence gauge at 0.9, banner Strands Decider

ALT Pencil sketch of a retro AI robot at a fork in the road choosing between option cards with a confidence gauge at 0.9, banner Strands Decider

12
Brittle UI test scripts are a QA tax. Amazon Bedrock AgentCore Browser + Amazon Nova Act turn JSON test cases into parallel, isolated cloud browser sessions with live view and session replay. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot running QA tests across parallel browser windows, with banner Agentic QA Automation and signs AgentCore Browser and Nova Act

ALT Pencil sketch of a retro AI robot running QA tests across parallel browser windows, with banner Agentic QA Automation and signs AgentCore Browser and Nova Act

23
Multi-turn agents fail when one early turn corrupts the rest. AEM scores agent quality per turn (Truthfulness + Completeness) to pinpoint the root-cause turn vs downstream effects. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot inspecting five numbered conversation-turn cards with a magnifying glass, circling turn 3 as the root-cause error with downstream cascade arrows, banner reading Agent Evaluation Metric, signs Truthfulness and Completeness

ALT Pencil sketch of a retro AI robot inspecting five numbered conversation-turn cards with a magnifying glass, circling turn 3 as the root-cause error with downstream cascade arrows, banner reading Agent Evaluation Metric, signs Truthfulness and Completeness

4
3
69
Model-agnostic PII detection with LLMs on Bedrock: detection logic lives in the prompt, so adding an entity type is a one-line edit — no retraining. Runs managed or self-hosted. Open-source. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot holding a magnifying glass over a form, highlighting PII fields, with a swappable model dial showing interchangeable model chips, banner reading Model-Agnostic PII Detection, signs Instruction-Driven and Any LLM on Bedrock

ALT Pencil sketch of a retro AI robot holding a magnifying glass over a form, highlighting PII fields, with a swappable model dial showing interchangeable model chips, banner reading Model-Agnostic PII Detection, signs Instruction-Driven and Any LLM on Bedrock

25
AWS Bedrock AgentCore now auto-optimizes agent system prompts: a reflector reads production traces, proposes a revised prompt, guardrails enforce no >20% growth + safety. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot editing a System Prompt scroll with a red pen while reading trace logs, an upward performance graph behind, banner reading Optimizing Agent Prompts, with signs AgentCore and Trace-Driven

ALT Pencil sketch of a retro AI robot editing a System Prompt scroll with a red pen while reading trace logs, an upward performance graph behind, banner reading Optimizing Agent Prompts, with signs AgentCore and Trace-Driven

21
Bedrock prompt caching: cache a static prefix (docs, prompts, tools) and reuse it — up to 90% lower cached-input cost, ~75% net savings on repeated context. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot filing a PROMPT document into a CACHE drawer with a stopwatch nearby, banner reading Prompt Caching, with signs 90% Lower Cost and Faster TTFT

ALT Pencil sketch of a retro AI robot filing a PROMPT document into a CACHE drawer with a stopwatch nearby, banner reading Prompt Caching, with signs 90% Lower Cost and Faster TTFT

17
Serverless PII redaction on AWS: Bedrock Data Automation flags sensitive fields by plain-language instruction, Step Functions + Lambda batch-redact. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot stamping black redaction bars over documents on a conveyor belt, banner reading Serverless PII Redaction, with signs Bedrock Data Automation and Step Functions

ALT Pencil sketch of a retro AI robot stamping black redaction bars over documents on a conveyor belt, banner reading Serverless PII Redaction, with signs Bedrock Data Automation and Step Functions

10
Picking a vector store for Amazon Bedrock Knowledge Bases? OpenSearch for low-latency hybrid search, Aurora pgvector when embeddings sit with relational data, S3 Vectors for large corpora at ~90% lower storage cost. Match the store to the use case. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot at a signpost with three arrows pointing to three crates of data vectors, banner reading Choosing a Vector Store, with signs OpenSearch, Aurora pgvector, and S3 Vectors

ALT Pencil sketch of a retro AI robot at a signpost with three arrows pointing to three crates of data vectors, banner reading Choosing a Vector Store, with signs OpenSearch, Aurora pgvector, and S3 Vectors

26
The new AWS AgentCore Runtime: memory reclaimed the instant a session ends (no more paying the peak) and consistent ~2s P75 cold starts regardless of image size. Serverless agents just got more elastic. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot sprinting on a track carrying a server container, banner reading AgentCore Runtime, with signs Fast Cold Starts and Scale to Zero

ALT Pencil sketch of a retro AI robot sprinting on a track carrying a server container, banner reading AgentCore Runtime, with signs Fast Cold Starts and Scale to Zero

17
Multi-tenant doc chat is easy to prototype, hard to operate. Amazon Bedrock Managed Knowledge Base owns ingestion, ranking & agentic multi-hop retrieval — enforcing per-tenant isolation from a verified identity on every hop. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot librarian at a MANAGED KB cart handing a TENANT C folder out, with tenants A-F each in their own locked booth; banner reads MULTI-TENANT DOC CHAT with ISOLATED and MANAGED KB signs

ALT Pencil sketch of a retro AI robot librarian at a MANAGED KB cart handing a TENANT C folder out, with tenants A-F each in their own locked booth; banner reads MULTI-TENANT DOC CHAT with ISOLATED and MANAGED KB signs

19
Want to give a talk at Boston Code Camp 41? Session submissions are open now! Go to bostoncodecamp.com/CC41/info and click the green "Offer Your Session" button. We look forward to your ideas!
2
2
65
As agentic AI scales, the problem shifts from building agents to governing them. AWS Agent Registry (now GA) is a single searchable, governed catalog for agents, tools & skills — publish once, discover and reuse across the enterprise. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot librarian filing labeled capability cards into a card catalog labeled REGISTRY, banner reading AWS AGENT REGISTRY with DISCOVER and GOVERN signs

ALT Pencil sketch of a retro AI robot librarian filing labeled capability cards into a card catalog labeled REGISTRY, banner reading AWS AGENT REGISTRY with DISCOVER and GOVERN signs

21
Running an agent in production is mostly ops, not reasoning. This walkthrough migrates a LangGraph support agent to Amazon Bedrock AgentCore in stages — Runtime, Gateway & Memory take the undifferentiated load off your plate. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot carrying a heavy OPS backpack up a staircase toward a clean cloud building, banner reading MIGRATE TO AGENTCORE with RUNTIME and MEMORY signposts

ALT Pencil sketch of a retro AI robot carrying a heavy OPS backpack up a staircase toward a clean cloud building, banner reading MIGRATE TO AGENTCORE with RUNTIME and MEMORY signposts

10
It’s official—Boston Code Camp 41 is November 21st, 2026. More details to come on the BCC website this September!
5
7
509
Three ways to cap Bedrock costs: CloudWatch alarms (5-15 min), AWS Budgets + tags (daily alerts), or an app-layer gateway (real-time per-user). Start with Budgets, escalate as needed. builder.aws.com/content/3DIY…
Pencil sketch of a retro AI robot watching three stacked budget dials labeled ALARM, BUDGET, and GATEWAY, each with a needle pointing to different levels, and a small dollar bill with a stop sign at the bottom. Banner reads LIMIT THE BILL, with labels BEDROCK and COST.

ALT Pencil sketch of a retro AI robot watching three stacked budget dials labeled ALARM, BUDGET, and GATEWAY, each with a needle pointing to different levels, and a small dollar bill with a stop sign at the bottom. Banner reads LIMIT THE BILL, with labels BEDROCK and COST.

30
AgentCore Gateway now supports rate limiting — token bucket algorithm, per-client and per-target throttling. One noisy agent can't starve the rest. No code changes, limit lives in gateway config. aws.amazon.com/blogs/machine…
Pencil sketch of a retro AI robot operating a valve on a pipe carrying flowing request arrows, with a gauge showing the rate limit level and a bucket below catching overflow drops. Banner reads RATE LIMITS, with labels AGENTCORE and GATEWAY.

ALT Pencil sketch of a retro AI robot operating a valve on a pipe carrying flowing request arrows, with a gauge showing the rate limit level and a bucket below catching overflow drops. Banner reads RATE LIMITS, with labels AGENTCORE and GATEWAY.

31
Load more