<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Julien Taste on Medium]]></title>
        <description><![CDATA[Stories by Julien Taste on Medium]]></description>
        <link>https://medium.com/@julien.taste?source=rss-10c592ba8d18------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*ORHQGRlvsPq1wv5oIxnw8w.jpeg</url>
            <title>Stories by Julien Taste on Medium</title>
            <link>https://medium.com/@julien.taste?source=rss-10c592ba8d18------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 21:20:20 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@julien.taste/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[How I Passed the AWS Certified Generative AI Developer — Professional (AIP-C01) Beta Exam with a…]]></title>
            <link>https://medium.com/@julien.taste/how-i-passed-the-aws-certified-generative-ai-developer-professional-aip-c01-beta-exam-with-a-bc7225872275?source=rss-10c592ba8d18------2</link>
            <guid isPermaLink="false">https://medium.com/p/bc7225872275</guid>
            <category><![CDATA[aws-certification]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Julien Taste]]></dc:creator>
            <pubDate>Mon, 23 Mar 2026 07:26:27 GMT</pubDate>
            <atom:updated>2026-03-23T07:26:27.385Z</atom:updated>
            <content:encoded><![CDATA[<h3>How I Passed the AWS Certified Generative AI Developer — Professional (AIP-C01) Beta Exam with a Score of 857/1000</h3><h3>A hands-on guide to one of AWS’s most challenging certifications in 2026</h3><p><em>By Julien Taste — Cloud &amp; AI Architect | March 2026</em></p><h3>Introduction</h3><p>On March 22, 2026, I sat for the beta version of the AWS Certified Generative AI Developer — Professional (AIP-C01) exam. Result: <strong>857/1000</strong>, with all five domains rated “Meets Competencies.”</p><p>This certification is the latest Professional-level credential from AWS, sitting alongside Solutions Architect Professional and DevOps Engineer Professional at the top of the AWS certification pyramid. It validates the ability to design, deploy, secure, and optimize production-grade generative AI applications on AWS.</p><p>Having worked extensively with Amazon Bedrock, RAG architectures, AI agents, and the Model Context Protocol (MCP) in enterprise environments, I wanted to formalize that hands-on expertise with an industry-recognized credential. This article shares my full experience: what the exam covers, how I prepared, the traps to watch for, and concrete tips for anyone considering this certification.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_GeQRJdxGfrlXoKw1IfJMA.jpeg" /></figure><h3>Why This Certification Matters</h3><p>Generative AI has moved beyond experimentation. Organizations are deploying production-grade GenAI applications at scale, and they need architects and developers who can make the right decisions under real constraints — cost, latency, security, compliance, and scalability.</p><p>The AIP-C01 tests exactly that. It doesn’t ask you to explain what a transformer is or how embeddings work. It asks you to choose the right AWS service for a specific use case, with multiple competing requirements. Every question is a mini architecture decision.</p><p>If you’re already building GenAI applications on AWS, this certification validates what you already know. If you’re transitioning into GenAI, it provides a structured learning path for production-ready skills.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mGo2rwEK0OaPJ0kqa6Pzjg.jpeg" /></figure><h3>What the Exam Covers</h3><p>The exam is organized into five domains:</p><p><strong>Domain 1 — Foundation Model Integration, Data Management &amp; Compliance (31%)</strong></p><p>The heaviest domain. Expect questions on RAG architectures with Bedrock Knowledge Bases, vector store selection (OpenSearch Serverless, Neptune Analytics for Graph RAG, MemoryDB, Aurora pgvector, S3 Vectors), chunking strategies (semantic, hierarchical, fixed-size), hybrid search with rerankers, metadata filtering, and data compliance patterns.</p><p>This domain alone represents nearly a third of the exam. If you master one domain, make it this one.</p><p><strong>Domain 2 — Implementation &amp; Integration (26%)</strong></p><p>The second largest block covers Bedrock Agents (action groups, Return of Control), AgentCore (memory, session, identity), multi-agent architectures (supervisor/collaborator patterns), Bedrock Flows for pipeline orchestration, MCP server deployment, Step Functions orchestration, the Converse API, streaming patterns, CI/CD deployment, and cross-environment integration.</p><p>The key skill here is knowing when to use Agents vs. Flows vs. Step Functions — each has a specific sweet spot.</p><p><strong>Domain 3 — AI Safety, Security &amp; Governance (20%)</strong></p><p>Bedrock Guardrails in depth: content filters, denied topics, PII filters (block vs. mask), contextual grounding checks, and the newer Automated Reasoning checks. Also covers VPC endpoints for private connectivity, IAM and SCPs for model access control, Lake Formation for column-level cross-account access, CloudFormation StackSets for multi-account guardrail deployment, and defense-in-depth patterns.</p><p><strong>Domain 4 — Operational Efficiency &amp; Optimization (12%)</strong></p><p>Provisioned throughput vs. on-demand, cross-region inference profiles for data residency and traffic distribution, prompt caching, semantic caching with OpenSearch k-NN, AWS AppConfig for dynamic configuration, intelligent prompt routing, and token management strategies.</p><p><strong>Domain 5 — Testing, Validation &amp; Troubleshooting (11%)</strong></p><p>Bedrock Model Evaluation jobs (retrieve-only vs. retrieve-and-generate), LLM-as-a-judge evaluation, SageMaker Clarify for fairness/bias detection, CloudWatch native metrics and anomaly detection, model invocation logging, and canary deployment strategies.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*V43b469kgKkoHLat5JSUTw.jpeg" /></figure><h3>My Background Going In</h3><p>I didn’t approach this exam from a purely theoretical standpoint. My daily work as a Cloud &amp; AI Architect involves:</p><ul><li>Designing and deploying RAG architectures with Bedrock Knowledge Bases</li><li>Building multi-agent systems with Bedrock Agents and AgentCore</li><li>Integrating MCP servers into enterprise workflows</li><li>Implementing GenAI security with Guardrails, VPC endpoints, and SCPs</li><li>Monitoring GenAI applications with CloudWatch and model invocation logging</li><li>Optimizing costs with cross-region inference and caching strategies</li></ul><p>This hands-on experience was my strongest asset. The exam rewards practical knowledge — understanding not just what a service does, but when and why to choose it over alternatives.</p><p>If you have similar production experience, you’re already well-positioned. If you’re newer to AWS GenAI services, I’d recommend spending significant time with the hands-on labs and the official documentation before attempting this exam.</p><h3>The Beta Exam Experience</h3><p>The beta version runs <strong>215 minutes</strong> for <strong>85 questions</strong> (65 scored + 10 unscored + 10 additional beta questions). That’s roughly 2 minutes 30 seconds per question.</p><p>Here’s what stood out:</p><p><strong>Questions are dense.</strong> Each question is a realistic architecture scenario with layered constraints — data residency AND cost optimization AND low latency AND security. You need to hold multiple requirements in mind simultaneously and find the answer that satisfies all of them.</p><p><strong>Answers are plausible.</strong> On most questions, 2–3 answers could work in theory. The difference often comes down to a single word in the question stem. “LEAST effort” eliminates custom solutions. “Defense-in-depth” requires multiple security layers. “Without manual intervention” disqualifies anything with “external logic” or “manually configure.”</p><p><strong>Time management matters.</strong> I used nearly all 215 minutes. If you’re spending more than 3 minutes on a question, flag it and move on. Come back with fresh eyes after completing the rest.</p><p><strong>The unscored questions are invisible.</strong> 10 of the 85 questions don’t count toward your score, but you can’t tell which ones. If you encounter a question that seems impossibly ambiguous, it might be an unscored experimental question. Don’t let it shake your confidence.</p><h3>10 Decision Patterns That Matter Most</h3><p>These are the mental models I used most frequently during the exam:</p><p><strong>1. “LEAST effort/overhead”</strong> → Always choose the native Bedrock service. If you’re choosing between Bedrock Knowledge Bases and a custom OpenSearch + Lambda pipeline, the native option wins.</p><p><strong>2. Predefined GenAI pipeline (S3 → parse → model → S3)</strong> → Bedrock Flows. Not Step Functions, not Agents.</p><p><strong>3. Conversational + external APIs + multi-turn</strong> → Bedrock Agents with action groups.</p><p><strong>4. Complex workflow + retry/catch + human approval</strong> → Step Functions Standard. Never Express for callbacks (Express has a 5-minute limit and no callback support).</p><p><strong>5. EU data residency</strong> → Cross-Region Inference profiles with eu.* geographic codes + SCPs for enforcement.</p><p><strong>6. Similar queries with semantic matching</strong> → OpenSearch k-NN for semantic caching. Bedrock native prompt caching for exact prompt matches only.</p><p><strong>7. Frontend streaming</strong> → InvokeModelWithResponseStream via HTTP API. Critical: REST API does NOT support Lambda response streaming.</p><p><strong>8. Reduce hallucinations</strong> → RAG + Guardrails contextual grounding with HIGH threshold. For mathematical/logical validation → Automated Reasoning checks (different safeguard, different purpose).</p><p><strong>9. “Defense-in-depth”</strong> → Minimum 2 security layers (e.g., Guardrails + WAF). A single security service is never defense-in-depth.</p><p><strong>10. Dynamic config without redeployment</strong> → AWS AppConfig (with feature flags, JSON schema validation, and automatic rollback). Not SSM Parameter Store (no rollback, no validation).</p><h3>5 Common Traps</h3><p><strong>Trap 1: REST API vs. HTTP API.</strong> Lambda response streaming only works with HTTP API. Any answer combining REST API with streaming is incorrect.</p><p><strong>Trap 2: “Denied topics” vs. “Topic policies.”</strong> The correct Bedrock Guardrails feature name is “denied topics.” “Topic policies” is a distractor designed to look similar.</p><p><strong>Trap 3: Provisioned throughput not being used.</strong> If code calls invoke_model(modelId=&quot;anthropic.claude-v2&quot;) with a string ID, it uses on-demand pricing even if provisioned throughput was purchased. The fix: replace the model ID string with the ARN returned by CreateProvisionedModelThroughput.</p><p><strong>Trap 4: Step Functions Express vs. Standard.</strong> Express: 5-minute maximum, no callbacks. Standard: up to 1 year, supports waitForTaskToken and human approval. When the question mentions “human review” or “callback,” it’s always Standard.</p><p><strong>Trap 5: retrieve-only vs. retrieve-and-generate evaluation.</strong> If the question asks to evaluate retrieval quality alone → retrieve-only. If it asks to evaluate the full RAG pipeline (retrieval + generation) → retrieve-and-generate. Mixing these up is an easy mistake.</p><h3>Key Service Associations</h3><p>Quick reference for the most tested service mappings:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*p5OpuMfnJ_3S0eAw8RLisw.png" /></figure><h3>Recommended Preparation Resources</h3><ol><li><strong>The official AWS exam guide</strong> — Read it thoroughly. Every skill listed maps to potential questions.</li><li><strong>AWS Bedrock documentation</strong> — Especially the pages on Knowledge Bases, Agents, Guardrails (all safeguard types), Flows node types, and Cross-Region Inference. This is the source of truth.</li><li><strong>Hands-on experience</strong> — Build a RAG application. Create an agent with action groups. Configure guardrails. Set up cross-region inference. There’s no substitute for having done it yourself.</li><li><strong>AWS Well-Architected Generative AI Lens</strong> — Several questions reference this framework explicitly. Understanding its principles will help with the governance and architecture questions.</li></ol><h3>My Results</h3><p><strong>Score: 857/1000</strong> (passing score: 750)</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n5u9fdI0ezyGV5d5XBznsw.jpeg" /></figure><h3>Final Thoughts</h3><p>The AIP-C01 is a demanding exam. It tests depth, not breadth. Every question requires you to evaluate trade-offs between multiple valid approaches and select the one that best satisfies all stated constraints.</p><p>My main advice: <strong>learn the decision patterns, not the documentation.</strong> The exam doesn’t test whether you can recite the Bedrock API reference. It tests whether you can look at a scenario with competing requirements and pick the right architecture.</p><p>When you see “LEAST effort,” think native. When you see “defense-in-depth,” count the layers. When you see “without manual intervention,” eliminate the custom options. These signals are your compass.</p><p>The beta exam closes on March 31, 2026. If you have production experience with AWS GenAI services and you’re reading this before that date, I’d encourage you to take the challenge. The standard version will follow later in 2026 for those who prefer a more stable exam.</p><p>Good luck to everyone preparing for this certification. The GenAI space on AWS is evolving fast, and this credential is a strong signal that you can keep up.</p><p><em>Julien Taste is a Cloud &amp; AI Architect with 20+ years of experience in cloud infrastructure and 10+ years in Cloud, AI, and DevSecOps. He currently works at Crédit Agricole Group Infrastructure Platform (CA-GIP) as a Cloud &amp; AI Architect via Horizon Consulting, based in France.</em></p><p><strong>Tags:</strong> AWS Generative AI Certification Amazon Bedrock Cloud Architecture AI Machine Learning RAG LLM AIP-C01</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bc7225872275" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Building a Real-Time Police Chase Analyzer with Gemini’s Full AI Stack]]></title>
            <link>https://medium.com/@julien.taste/building-a-real-time-police-chase-analyzer-with-geminis-full-ai-stack-de701ee7c0e7?source=rss-10c592ba8d18------2</link>
            <guid isPermaLink="false">https://medium.com/p/de701ee7c0e7</guid>
            <dc:creator><![CDATA[Julien Taste]]></dc:creator>
            <pubDate>Mon, 16 Mar 2026 19:40:18 GMT</pubDate>
            <atom:updated>2026-03-16T19:40:18.236Z</atom:updated>
            <content:encoded><![CDATA[<h3>How I built Pursuit Analyst — a multimodal AI agent that watches, understands, and talks about police pursuits in real time</h3><p>Every year, tens of thousands of police pursuits unfold across the United States. Each one is a high-stakes, fast-moving event where critical decisions happen in seconds — speeds escalate, suspects weave through traffic, and officers must balance public safety against the need to apprehend. Analyzing these events after the fact typically requires hours of manual video review.</p><p>I built <strong>Pursuit Analyst</strong> to compress that process into 40 seconds using the full breadth of Google’s Gemini ecosystem — and then let you have a live voice conversation with an AI that knows every detail of what happened.</p><p>This article walks through the architecture, the technical challenges, and the lessons learned while building it for the <strong>Gemini Live Agent Challenge</strong>.</p><h3>What Pursuit Analyst Does</h3><p>Give it a YouTube URL of a police chase. Click one button. In under a minute, you get:</p><ul><li><strong>Semantic tactical phases</strong> — every distinct moment of the pursuit identified, timestamped, and scored by danger level (1–5)</li><li><strong>Suspect vehicle identification</strong> — make, model, year, color, plate visibility, damage assessment, plus an AI-generated reference image</li><li><strong>High-frequency tracking</strong> — 50 data points mapping suspect and police speeds, distances, and environment changes</li><li><strong>Dual vector embeddings</strong> — every phase indexed as both a text embedding and a multimodal video embedding, enabling semantic search across what was <em>described</em> and what was <em>visually seen</em></li><li><strong>Pre-cached audio commentary</strong> — AI-generated voice analysis for every phase, ready to play in sync with the video</li><li><strong>AI-generated summary video</strong> — an 8-second cinematic reconstruction of the most dangerous moments</li></ul><p>And then you can talk to it. Out loud. In real time. Ask it what happened at minute 3. Ask it to describe the suspect. Ask it about the most dangerous moment. It answers instantly with full context.</p><h3>The Architecture: 7 Gemini APIs Working Together</h3><p>The entire application runs on Google Cloud with zero external APIs. Here’s the stack:</p><h4>1. Gemini 3 Flash — The Brain</h4><p>The analysis pipeline starts with gemini-3-flash-preview performing three tasks in parallel:</p><p><strong>Semantic Chunking</strong> — The video is analyzed and broken into structured tactical phases. Each phase is a JSON object with start/end timestamps, maneuver type, estimated speed, danger level, environment description, and a keyword. A 10-minute pursuit typically yields 10–12 phases.</p><p><strong>Suspect Tracking</strong> — 50 high-frequency tracking points are generated, each containing suspect speed, police speed, distance estimate, environment, and tactical notes.</p><p><strong>Suspect Identification</strong> — Individual frames are extracted and analyzed to identify the vehicle: make, model, year, color, body type, plate visibility, and damage.</p><p>All three run concurrently via asyncio, and the whole step completes in about 23 seconds for a typical 10-minute video.</p><h4>2. Gemini Embedding — Dual Vector Indexing</h4><p>This is where it gets interesting. Each tactical phase is embedded <em>twice</em>:</p><p><strong>Text embeddings</strong> — The phase description (maneuver, speed, danger, environment) is embedded via gemini-embedding-2-preview into a 768-dimensional vector.</p><p><strong>Video embeddings</strong> — The actual video segment is split with ffmpeg, uploaded to the Gemini Files API, and embedded as a <em>multimodal</em> vector combining the visual content with its text context.</p><p>These two sets of vectors go into two separate <strong>Vertex AI Vector Search</strong> indexes. This means you can search by description (“sudden braking”) <em>and</em> by visual similarity — two completely different retrieval dimensions over the same data.</p><p>Why not one index? Because text and video embeddings capture fundamentally different information. A phase described as “highway straight line” might visually show a near-miss that the text doesn’t capture. Dual indexes let you find both.</p><pre># Text embedding<br>result = client.models.embed_content(<br>    model=&quot;gemini-embedding-2-preview&quot;,<br>    contents=phase_description,<br>    config=types.EmbedContentConfig(<br>        task_type=&quot;RETRIEVAL_DOCUMENT&quot;,<br>        output_dimensionality=768,<br>    ),<br>)</pre><pre># Video embedding (multimodal)<br>video_file = client.files.upload(file=segment_path)<br>result = client.models.embed_content(<br>    model=&quot;gemini-embedding-2-preview&quot;,<br>    contents=[video_file, context_text],<br>    config=types.EmbedContentConfig(<br>        task_type=&quot;RETRIEVAL_DOCUMENT&quot;,<br>        output_dimensionality=768,<br>    ),<br>)</pre><h4>3. Vertex AI Vector Search — Production-Grade Retrieval</h4><p>Both indexes use STREAM_UPDATE mode for real-time upserts, cosine distance, and token-based restricts for filtering by phase type, maneuver, or minimum danger level.</p><p>One important lesson: <strong>Vertex AI Vector Search doesn’t store rich metadata</strong>. It only stores vectors and simple restricts. So I store the full metadata (timestamps, descriptions, danger scores) as JSON files on Google Cloud Storage and load them at startup. The search returns vector IDs, which I map back to the GCS metadata.</p><pre># Upsert with token restricts for filtering<br>datapoint = IndexDatapoint(<br>    datapoint_id=f&quot;text_{video_id}_{i}&quot;,<br>    feature_vector=embedding,<br>    restricts=[<br>        IndexDatapoint.Restriction(namespace=&quot;type&quot;, allow_list=[phase_type]),<br>        IndexDatapoint.Restriction(namespace=&quot;danger_gte&quot;, allow_list=[&quot;3&quot;]),<br>    ],<br>)</pre><h4>4. Imagen 4 — Suspect Visualization</h4><p>Once Gemini Flash identifies the suspect vehicle, Imagen 4 generates a reference image from the description. This gives investigators a clean visual reference even when the helicopter footage is grainy or at a bad angle.</p><h4>5. Gemini Native Audio — Two Roles</h4><p>Native audio serves two purposes in Pursuit Analyst:</p><p><strong>TTS Pre-Cache</strong> — Before the user even starts watching, all phase commentaries are generated as PCM audio using gemini-2.5-flash-native-audio-preview with the Orus voice. The audio chunks are sent to the frontend via WebSocket, and JavaScript plays them at the exact video timestamp. Result: zero latency during playback.</p><p><strong>Live Voice Agent</strong> — This is the centerpiece. A full bidirectional streaming voice agent built with the Google Agent Development Kit (ADK).</p><h4>6. Google ADK — The Live Voice Agent</h4><p>The voice agent is built using ADK’s LiveRequestQueue and run_live() for real-time bidirectional audio streaming. Here&#39;s what makes it work:</p><p><strong>Context injection</strong> — At session start, the agent’s instruction is populated with the pursuit analysis: top 5 tactical phases by danger, suspect details, and sampled tracking data. The agent doesn’t need to re-analyze anything — it already knows the key details.</p><p><strong>Function tools</strong> — The agent has access to get_moment_details() and get_tracking_at_time(). It can look up specific timestamps and provide precise speed/distance data on demand.</p><p><strong>Audio pipeline</strong> — The browser captures microphone audio via ScriptProcessor (1024 samples, ~23ms chunks), resamples from native rate to 16kHz, filters silence, and streams Base64 PCM over WebSocket. The server feeds it to ADK as Blob(audio/pcm;rate=16000). ADK returns 24kHz PCM audio that&#39;s streamed back and played with lookahead scheduling for smooth output.</p><pre>voice_agent = Agent(<br>    name=&quot;pursuit_voice&quot;,<br>    model=&quot;gemini-2.5-flash-native-audio-preview-12-2025&quot;,<br>    instruction=f&quot;&quot;&quot;Pursuit analyst dispatcher. 1-2 sentences max.<br>    SUSPECT: {suspect_desc}<br>    PHASES: {phases_summary}<br>    TRACKING: {tracking_summary}&quot;&quot;&quot;,<br>    tools=[<br>        FunctionTool(get_moment_details),<br>        FunctionTool(get_tracking_at_time),<br>    ],<br>)</pre><p>The key insight: <strong>keep the instruction concise</strong>. The model processes the full instruction on every turn. A 500-token instruction adds measurable latency compared to a 150-token one. I moved detailed data into the tools rather than cramming everything into the system prompt.</p><h4>7. Veo 3.1 — AI Video Summary</h4><p>The final feature: click a button, and Veo 3.1 generates an 8-second cinematic reconstruction of the pursuit’s most dangerous moments.</p><p>The prompt is built programmatically from the top 5 danger-scored phases, including maneuvers, speeds, and environment details. The result is a surveillance-style aerial video that captures the intensity of the chase.</p><pre>operation = client.models.generate_videos(<br>    model=&quot;veo-3.1-fast-generate-preview&quot;,<br>    prompt=cinematic_prompt,<br>    config=types.GenerateVideosConfig(<br>        aspect_ratio=&quot;16:9&quot;,<br>        resolution=&quot;720p&quot;,<br>    ),<br>)</pre><pre># Poll until done (1-3 minutes)<br>while not operation.done:<br>    time.sleep(10)<br>    operation = client.operations.get(operation)</pre><pre># Download the result<br>video = operation.response.generated_videos[0]<br>client.files.download(file=video.video)<br>video.video.save(tmp_path)</pre><p>One gotcha: you <strong>must</strong> call client.files.download() before accessing the video bytes. The initial response only contains a reference — the actual bytes require a separate download step.</p><h3>The Frontend: A Single HTML File</h3><p>The entire frontend is one HTML file. No build tools, no framework, no npm. Just vanilla JavaScript with WebSocket connections and the Web Audio API.</p><p>Two WebSocket endpoints handle everything:</p><ul><li>/ws/chat — Analysis pipeline, text chat, vector search, Veo generation</li><li>/ws/voice — Live audio streaming with the ADK agent</li></ul><p>The audio playback uses a lookahead scheduling pattern:</p><pre>const startAt = Math.max(now + 0.005, _voiceNextTime);<br>src.start(startAt);<br>_voiceNextTime = startAt + buf.duration;<br>const delay = Math.max(10, (_voiceNextTime - now - 0.05) * 1000);<br>setTimeout(scheduleVoiceBuffer, delay);</pre><p>This ensures gapless audio output while keeping latency minimal. The 50ms lookahead and 5ms start delay were tuned through experimentation — larger values sound smoother but feel sluggish.</p><h3>Deployment: Cloud Run with WebSocket Persistence</h3><p>The app runs on Google Cloud Run with a few critical flags:</p><ul><li>--no-cpu-throttling — Without this, Cloud Run throttles CPU when there&#39;s no active request. But WebSocket connections are long-lived, and background tasks (video embeddings, Veo generation) need CPU even when the client isn&#39;t sending messages.</li><li>--session-affinity — Ensures the same client hits the same instance. Critical because the analysis state is in memory.</li><li>--timeout 900 — Voice sessions and Veo generation can run for minutes.</li></ul><p>Cost control was a real concern. The mitigations:</p><ul><li>IP whitelist via ALLOWED_IPS environment variable</li><li>Rate limiting: one analysis per WebSocket session</li><li>Max 5 concurrent connections</li><li>IAM-based access toggle for quick enable/disable</li></ul><h3>Lessons Learned</h3><p><strong>1. Vertex AI Vector Search metadata is minimal.</strong> Plan for external metadata storage from the start. I used GCS JSON files, which works but adds a loading step at startup.</p><p><strong>2. Gemini Files API requires polling.</strong> Uploaded files go through a PROCESSING state before becoming ACTIVE. Budget 2 minutes per video segment upload + processing.</p><p><strong>3. Native audio context windows fill fast.</strong> At ~25 tokens per second of audio, a 5-minute conversation consumes significant context. Keep sessions focused or they will degrade.</p><p><strong>4. Keep voice agent instructions short.</strong> Every token in the system instruction is processed on every turn. Move detailed data into tools that the model can call on demand.</p><p><strong>5. Cloud Run and WebSockets need special care.</strong> Without --no-cpu-throttling and --session-affinity, long-running WebSocket connections will break randomly.</p><p><strong>6. Veo requires explicit file download.</strong> The generate_videos response contains a reference, not bytes. Call client.files.download() then video.save() to get the actual video.</p><p><strong>7. Dual embeddings are worth the complexity.</strong> Text search finds what’s described. Video search finds what’s shown. The overlap is surprisingly small — about 60% of searches return different top results from each index.</p><h3>The Numbers</h3><ul><li><strong>40 seconds</strong> for complete analysis of a 10-minute pursuit</li><li><strong>11 tactical phases</strong> identified on average</li><li><strong>50 tracking points</strong> per video</li><li><strong>768-dimensional</strong> dual vector embeddings</li><li><strong>2 Vertex AI indexes</strong> (text + video)</li><li><strong>7 Google APIs</strong> working together</li><li><strong>1 HTML file</strong> for the entire frontend</li><li><strong>0 external APIs</strong> — 100% Google Cloud</li></ul><h3>Try It / Learn More</h3><p>Pursuit Analyst was built for the Gemini Live Agent Challenge 2026.</p><p><strong>GitHub</strong>: <a href="https://github.com/julien2613/pursuit-analyst">github.com/julien2613/pursuit-analyst</a></p><p><strong>Stack</strong>: Gemini 3 Flash, Gemini Embedding, Gemini Native Audio, Imagen 4, Veo 3.1, Vertex AI Vector Search, Google ADK, Cloud Run, Cloud Storage, Secret Manager.</p><p>This article was created for the purposes of entering the <strong>Gemini Live Agent Challenge 2026</strong> hackathon on Devpost. <strong>#GeminiLiveAgentChallenge</strong></p><p><em>Built with the Google Agent Development Kit and the Gemini API. All infrastructure on Google Cloud.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=de701ee7c0e7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Beyond Skills vs MCP: Dynamic Agent Orchestration for Real-World Enterprise AI]]></title>
            <link>https://medium.com/@julien.taste/beyond-skills-vs-mcp-dynamic-agent-orchestration-for-real-world-enterprise-ai-908fa0a98cb4?source=rss-10c592ba8d18------2</link>
            <guid isPermaLink="false">https://medium.com/p/908fa0a98cb4</guid>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[entrepreneurship]]></category>
            <category><![CDATA[ai-agents-in-action]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[software-development]]></category>
            <dc:creator><![CDATA[Julien Taste]]></dc:creator>
            <pubDate>Sat, 27 Dec 2025 07:51:14 GMT</pubDate>
            <atom:updated>2025-12-27T07:57:28.488Z</atom:updated>
            <content:encoded><![CDATA[<p>Mohamed Elkholy’s recent article on Claude Code capabilities presents a clean mental model: MCP provides access, Skills provide expertise.</p><p>While this framework works for simple workflows, enterprise AI deployments require something more sophisticated — agents that don’t just follow playbooks but dynamically orchestrate tools based on evolving context.</p><p>After 12 months building AI prototypes</p><p>I’ve found that the <strong>MCP/Skills dichotomy</strong> is a starting point, not a destination. Here’s what actually works in production.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fmn3d3jJN050%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dmn3d3jJN050&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fmn3d3jJN050%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/cf2c49ade60772bf4ff6dd94418dbbcc/href">https://medium.com/media/cf2c49ade60772bf4ff6dd94418dbbcc/href</a></iframe><h3>The Three Maturity Levels of Agent Orchestration</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B4bexzakvSs6hwc5uZTtGg.png" /><figcaption><em>The three levels of agent sophistication: Static → Dynamic → Multi-Agent</em></figcaption></figure><h3>Level 1: Static Orchestration (The Elkholy Model)</h3><p>This is where most teams start. MCP servers expose tools (CRM, databases, APIs), and Skills define fixed sequences: “Step 1: query customer data → Step 2: check inventory → Step 3: format proposal.”</p><p><strong>This works when:</strong></p><ul><li>Workflows are predictable and linear</li><li>Edge cases are minimal</li><li>Output format is standardized</li></ul><p><strong>The limitation?</strong> Real business conversations don’t follow scripts. A sales agent that rigidly follows Step 1–2–3 will fail when a customer asks an unexpected question at Step 2.</p><h3>Level 2: Context-Aware Dynamic Skills</h3><p>The shift happens when Skills stop being playbooks and start being <strong>decision frameworks</strong>. Instead of “do A then B then C,” the agent evaluates conversational state and business context before choosing which MCP to invoke, in what order, with what parameters.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-i9ebqHveBGVhUozaYf6cA.png" /><figcaption><em>Static skills break on interruptions. Dynamic skills adapt and re-prioritize.</em></figcaption></figure><p>For an AI commercial agent handling advertising sales, this means:</p><ul><li>The agent <strong>assesses buyer intent signals</strong> before deciding whether to pull pricing data or case studies first</li><li>It <strong>adapts its MCP query strategy</strong> based on what information is already in context</li><li>It <strong>handles interruptions and topic changes</strong> without losing thread</li></ul><p><strong>The key architectural change:</strong> Skills become stateful reasoning loops, not linear sequences.</p><h3>Level 3: Multi-Agent Collaboration</h3><p>The most sophisticated deployments use specialized agents that delegate to each other. A “sales orchestrator” might invoke:</p><ul><li>A <strong>pricing specialist</strong> agent (with its own MCP connections to rate cards and market data)</li><li>A <strong>content specialist</strong> agent (connected to campaign archives and creative assets)</li></ul><p>Each sub-agent has its own Skills, but the orchestrator decides when to engage which specialist based on conversation flow.</p><h3>Practical Implementation: What We Learned</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lVCOWPiqaX1gMY13vf-95g.png" /><figcaption><em>Real-world implementation: Context-aware orchestration for advertising sales</em></figcaption></figure><h3>Context Window as Strategic Asset</h3><p>MCP calls consume tokens. In long sales conversations, naive implementations hit context limits fast. We implemented a <strong>“context compression” layer</strong> in our Skills that summarizes MCP responses before storing them in conversational memory — keeping signal, discarding noise.</p><h3>Graceful Degradation</h3><p>When MCP servers timeout or return errors (and they will), Skills need fallback logic. Our agents maintain <strong>“last known good” states</strong> and can continue conversations with cached data while flagging uncertainty to users.</p><h3>Human-in-the-Loop Checkpoints</h3><p>For high-stakes decisions (pricing approvals, contract terms), Skills include <strong>explicit handoff points</strong>. The agent prepares the decision context, but a human validates before the agent proceeds.</p><h3>The Tools We Use</h3><p>Our stack combines several frameworks depending on the use case:</p><p>Framework :</p><p><strong>AWS Bedrock AgentCore : </strong>Enterprise deployments with compliance requirements</p><p><strong>Google ADK : </strong>Flexibility for rapid prototyping</p><p><strong>CopilotKit : </strong>Tight frontend integration for user-facing agents</p><p>The common thread: all these frameworks support the <strong>dynamic orchestration pattern</strong> — Skills that reason about context before acting.</p><h3>When to Use What: A Revised Decision Framework</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YoLkoQQUd2mz-hcQbmyWpQ.png" /><figcaption><em>Match your architecture to your use case complexity</em></figcaption></figure><p>Problem You Need to Solve</p><p>Simple data retrieval, fixed format : <strong>MCP alone</strong></p><p>Predictable multi-step workflows : <strong>Static Skills + MCP</strong></p><p>Conversational agents with variable paths : <strong>Dynamic Skills + MCP</strong></p><p>omplex enterprise processes with multiple domains : <strong>Multi-Agent + Dynamic Skills + MCP</strong></p><h3>Conclusion</h3><p>The MCP/Skills distinction is pedagogically useful but practically incomplete. Real enterprise AI requires agents that treat Skills as <strong>reasoning frameworks</strong>, not execution scripts.</p><p>The tools are mature enough. The challenge is architectural: designing agents that balance <strong>structure</strong> (for reliability) with <strong>flexibility</strong> (for real-world messiness).</p><blockquote><strong><em>MCP gives you access. Static Skills give you consistency. Dynamic orchestration gives you intelligence.</em></strong></blockquote><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=908fa0a98cb4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Detection of Deviant Systems in Cybersecurity Using Bernoulli Convolutions and Multifractal…]]></title>
            <link>https://medium.com/@julien.taste/detection-of-deviant-systems-in-cybersecurity-using-bernoulli-convolutions-and-multifractal-2833f4d3c246?source=rss-10c592ba8d18------2</link>
            <guid isPermaLink="false">https://medium.com/p/2833f4d3c246</guid>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[cloud]]></category>
            <category><![CDATA[mathematics]]></category>
            <category><![CDATA[physics]]></category>
            <dc:creator><![CDATA[Julien Taste]]></dc:creator>
            <pubDate>Fri, 24 Jan 2025 06:13:26 GMT</pubDate>
            <atom:updated>2025-01-24T06:13:26.350Z</atom:updated>
            <content:encoded><![CDATA[<h3>Detection of Deviant Systems in Cybersecurity Using Bernoulli Convolutions and Multifractal Analysis</h3><p>We have developed an <strong>anomaly detection algorithm for identifying deviant systems in cybersecurity</strong>, leveraging <strong>Bernoulli convolutions</strong> and their <strong>Gibbs properties</strong>, which facilitate multifractal analysis. Our goal is to detect <strong>systems exhibiting abnormal behaviors</strong>, potentially indicating security breaches, misconfigurations, or emerging threats.</p><h3>1. Understanding Deviant Systems in Cybersecurity</h3><p>A <strong>deviant system</strong> is one that deviates from its expected behavior, often due to malicious activities, hardware failures, or operational misconfigurations. Such deviations can manifest as:</p><ul><li><strong>Unusual user activity</strong>, such as logins from unexpected locations or times,</li><li><strong>Irregular resource usage</strong>, including spikes in CPU, memory, or disk activity,</li><li><strong>Anomalous network traffic</strong>, such as unexpected data transfers or communication with unknown IPs,</li><li><strong>Unexpected system errors</strong>, which may indicate hidden malware or attacks in progress.</li></ul><p>Detecting these deviations early is crucial to <strong>preventing security breaches and minimizing operational disruptions.</strong> Our algorithm focuses on identifying these deviant behaviors using a mathematical approach that captures both short-term and long-term anomalies.</p><h3>2. Simulating Deviant System Behavior with Bernoulli Distribution</h3><p>To effectively detect anomalies, we simulate <strong>deviant system behaviors</strong> using a <strong>Bernoulli distribution</strong>, where system events are represented as binary values (0 for normal and 1 for abnormal occurrences). This allows us to model various scenarios such as:</p><ul><li>Normal user activity interspersed with occasional failed login attempts,</li><li>Gradual increases in failed access attempts leading to a potential brute-force attack,</li><li>Random but frequent system errors signaling underlying issues.</li></ul><p><strong>Why do we use this approach?</strong><br>By generating synthetic data, we can study how normal systems behave and how deviations emerge over time, allowing us to fine-tune our detection mechanism.</p><h3>3. Applying Bernoulli Convolution to Identify Deviant Patterns</h3><p>We apply <strong>Bernoulli convolutions</strong> to system event data, which allows us to analyze historical trends and detect <strong>gradual deviations from normal behavior.</strong> This approach helps in:</p><ul><li>Capturing <strong>persistent anomalies</strong> that develop over time, such as an increase in failed logins,</li><li>Detecting <strong>sudden spikes</strong> in abnormal events, indicating potential attacks or failures,</li><li>Differentiating between random noise and meaningful deviations.</li></ul><p><strong>Why do we use convolutions?</strong><br>Traditional rule-based detection systems may fail to recognize patterns hidden in a system’s history. Bernoulli convolutions provide a way to account for both immediate and past behaviors, offering a more robust detection capability.</p><h3>4. Statistical Analysis for Threshold-Based Detection</h3><p>Once the convolution is applied, we use <strong>statistical techniques</strong> to set thresholds for detecting deviations. By normalizing the convoluted data and analyzing deviations based on <strong>standard deviation</strong>, we can identify anomalies such as:</p><ul><li><strong>Sustained deviations</strong>, which may indicate ongoing attacks or chronic issues,</li><li><strong>Transient anomalies</strong>, such as temporary load spikes or system misconfigurations,</li><li><strong>Rare but critical deviations</strong>, which may signal zero-day attacks or insider threats.</li></ul><p><strong>Why is statistical analysis important?</strong><br>By quantifying deviations mathematically, we can ensure that detection is objective and minimize false positives while capturing critical incidents.</p><h3>5. Generating and Analyzing Synthetic Logs for Testing</h3><p>To validate our approach, we generate <strong>synthetic logs</strong> that mimic real-world cybersecurity incidents. These logs include events such as:</p><ul><li>Unauthorized access attempts with increasing frequency,</li><li>Repeated system alerts related to resource exhaustion,</li><li>Gradual performance degradation leading to potential denial of service.</li></ul><p><strong>Why do we generate synthetic logs?</strong><br>This allows us to test the effectiveness of our detection algorithm under controlled conditions, ensuring it can correctly identify deviant system behaviors.</p><h3>6. Visualization for Effective Monitoring and Response</h3><p>Once anomalies are detected, we visualize them through intuitive graphs that allow security analysts to:</p><ul><li>Pinpoint periods of abnormal activity,</li><li>Understand the progression of a deviant system over time,</li><li>Correlate detected anomalies with log entries to gain deeper insights.</li></ul><p><strong>Why is visualization key?</strong><br>Graphical representation of system deviations helps security teams take swift action and prevent further incidents.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/935/1*rS3KTad1woihw9r63eZ90Q.png" /><figcaption>Anomaly Detection Using Bernoulli Convolution</figcaption></figure><h3>Conclusion and Future Prospects</h3><p>By using Bernoulli convolutions and multifractal analysis, we have created a powerful framework for <strong>detecting deviant systems in cybersecurity.</strong> This approach allows us to:</p><ul><li><strong>Identify complex, evolving threats</strong> that evade traditional detection methods,</li><li><strong>Reduce false positives</strong> by distinguishing normal operational fluctuations from real issues,</li><li><strong>Provide scalable, adaptable solutions</strong> for different IT environments, whether cloud-based or on-premise.</li></ul><p>Looking ahead, we plan to enhance our system by incorporating <strong>machine learning models</strong> to further refine our detection capabilities and adapt to evolving cyber threats dynamically.</p><p>With our approach, organizations can <strong>proactively detect and mitigate deviant system behaviors,</strong> ensuring enhanced security and operational efficiency.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2833f4d3c246" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>