<?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 Saravanan on Medium]]></title>
        <description><![CDATA[Stories by Saravanan on Medium]]></description>
        <link>https://medium.com/@saru2020?source=rss-15e0a997f395------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*7h1erguwBIUbNxz_TPotnQ.png</url>
            <title>Stories by Saravanan on Medium</title>
            <link>https://medium.com/@saru2020?source=rss-15e0a997f395------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Tue, 17 Nov 2026 08:43:07 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@saru2020/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[Pocket TTS Demo: I cloned Rajnikanth’s voice on my machine, kinda!]]></title>
            <link>https://saru2020.medium.com/pocket-tts-demo-i-cloned-rajnikanths-voice-on-my-machine-kinda-4d59c1993d0e?source=rss-15e0a997f395------2</link>
            <guid isPermaLink="false">https://medium.com/p/4d59c1993d0e</guid>
            <category><![CDATA[text-to-speech]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[voice-cloning]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Saravanan]]></dc:creator>
            <pubDate>Sat, 17 Jan 2026 10:15:07 GMT</pubDate>
            <atom:updated>2026-01-17T10:43:12.020Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*67xszwNQQKNSqC_UQr2I5Q.png" /><figcaption>pocket_tts demo with voice cloning</figcaption></figure><p>Pocket TTS Demo: I cloned Rajnikanth’s voice on my machine, kinda!</p><p>I tried <a href="https://github.com/kyutai-labs/pocket-tts">pocket-tts</a> from Kyutai Labs and, with a short audio clip, ended up with a voice that sounded oddly familiar. No GPU setup, no cloud stack, just with my Mac.</p><p>Pocket TTS is a lightweight, CPU-friendly TTS demo. It can speak with a built-in voice or clone a voice from a short audio clip.</p><p><strong>Demo Repository:</strong> <a href="https://github.com/saru2020/AI_POCs/blob/main/Pocket_TTS">Check out the full demo on GitHub</a></p><p>The repo stays lean: one entry script ( src/pocket_tts_demo.py), a Docker runner ( run.sh), and helpers that keep the flow smooth. Both paths lead to the same moment: your text turns into a WAV file under output/.</p><h3>The quick path</h3><p>If you like containers, the flow is:</p><pre>./run.sh build ./run.sh run --voice alba --text &quot;Hello, this is a test&quot; --output /app/output/test.wav</pre><p>If you prefer Python:</p><pre>pip install pocket-tts --no-deps pip install -r requirements.txt python src/pocket_tts_demo.py --voice alba --text &quot;Hello world&quot; --output output/test.wav</pre><p>You can play the audio afterwards (macOS example):</p><pre>afplay output/test.wav</pre><h3>Voice cloning in three small steps</h3><p>Voice cloning is optional and gated by the model access rules. The repo already includes a demo input file ( input/rajni_audio.wav) so you can try it quickly.</p><ol><li><strong>Accept the model terms</strong></li><li><strong>Get a Hugging Face token</strong></li><li><strong>Run the demo with your audio</strong></li></ol><ul><li>With Docker:</li></ul><pre>HUGGINGFACE_HUB_TOKEN=your_token ./run.sh run \ --audio-file /app/input/rajni_audio.wav \ --text &quot;You know what: Anger! Anger is the cause of all miseries, one should know how to control it; otherwise life becomes miserable; and hey, last but not least&quot; \ --output /app/output/test.wav</pre><ul><li>With Python:</li></ul><pre>python src/pocket_tts_demo.py \ --audio-file input/rajni_audio.wav \ --text &quot;You know what: Anger! Anger is the cause of all miseries, one should know how to control it; otherwise life becomes miserable; and hey, last but not least&quot; \ --output output/test.wav</pre><p>That is it. The script loads the model, extracts a voice state from your audio file, and generates speech from your text. The output is a WAV you can play or share.</p><h3>A small ending</h3><p>The best part of this demo is how compact it feels. A short audio clip in input/, a line of text, and a single command later, you have a voice that sounds familiar reading something new. If you want to skim the mechanics, start with <a href="https://github.com/saru2020/AI_POCs/blob/main/Pocket_TTS/README.md">Pocket_TTS/README.md</a>. If you want to hear it, just play/run the audio/demo.</p><p><em>Originally published at </em><a href="https://saru2020.blogspot.com/2026/01/pocket-tts-demo-i-cloned-rajnikanths.html"><em>https://saru2020.blogspot.com</em></a><em> on January 17, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4d59c1993d0e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Fine-tuning an LLM]]></title>
            <link>https://saru2020.medium.com/fine-tuning-an-llm-69cc86cb3cc2?source=rss-15e0a997f395------2</link>
            <guid isPermaLink="false">https://medium.com/p/69cc86cb3cc2</guid>
            <category><![CDATA[fine-tuning]]></category>
            <category><![CDATA[large-language-models]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[machine-learning]]></category>
            <dc:creator><![CDATA[Saravanan]]></dc:creator>
            <pubDate>Fri, 19 Dec 2025 09:50:24 GMT</pubDate>
            <atom:updated>2025-12-19T10:04:01.674Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*bzvwkyxWhuBI_OJ0IGurMQ.png" /></figure><p>I recently fine-tuned TinyLlama on a small custom dataset and was impressed by how well it learned the specific response style. Here’s what I did and the results.</p><p>You can try it out yourself by checking out the <a href="https://github.com/saru2020/tinyllama-llm-finetuned.git">repository</a>.</p><h3>What is Fine-tuning?</h3><p>Fine-tuning takes a pre-trained language model (one that already understands general language) and trains it further on specific data to improve performance on particular tasks. Think of it as giving a general-purpose assistant specialized training in a specific domain.</p><h3>The Training Data</h3><p>I started with just 3 examples in a simple JSON format:</p><pre>[ {&quot;prompt&quot;: &quot;Explain Python lists&quot;, &quot;response&quot;: &quot;Python lists are ordered, mutable collections.&quot;}, {&quot;prompt&quot;: &quot;What is a dictionary?&quot;, &quot;response&quot;: &quot;A dictionary stores key-value pairs with fast lookup.&quot;}, {&quot;prompt&quot;: &quot;Explain list comprehension&quot;, &quot;response&quot;: &quot;A concise syntax for creating lists from iterables.&quot;} ]</pre><p>The goal was to teach the model to give concise, technical answers about Python concepts.</p><h3>The Setup</h3><p>I used <strong>LoRA (Low-Rank Adaptation)</strong> for parameter-efficient fine-tuning. Instead of updating all model parameters (which would require massive resources), LoRA only trains a small set of additional parameters-reducing memory usage by ~90% while maintaining nearly the same performance.</p><p>The training ran for 10 epochs on a CPU, taking about 2 minutes. The loss decreased from 3.03 to 0.18, showing clear learning progress.</p><h3>The Results</h3><p>After training, I compared the original model with the fine-tuned version:</p><h3>Example 1: “Explain Python lists”</h3><p><strong>Original Model:</strong></p><blockquote><em>A Python list is an ordered collection of objects.</em></blockquote><p><strong>Fine-tuned Model:</strong></p><blockquote><em>Python lists are ordered, mutable collections.</em></blockquote><p><strong>Expected (from training data):</strong></p><blockquote><em>Python lists are ordered, mutable collections.</em></blockquote><p><strong>Perfect match!</strong> The fine-tuned model produced the exact response from the training data.</p><h3>Example 2: “What is a dictionary?”</h3><p><strong>Original Model:</strong></p><blockquote><em>A dictionary is a collection of words, their definitions, and their meanings.</em></blockquote><p><strong>Fine-tuned Model:</strong></p><blockquote><em>A dictionary is a collection of key-value pairs.</em></blockquote><p><strong>Expected (from training data):</strong></p><blockquote><em>A dictionary stores key-value pairs with fast lookup.</em></blockquote><p>The fine-tuned model correctly interpreted “dictionary” as a Python data structure (not a book), though it didn’t match the exact wording. Still, it captured the core concept.</p><h3>Key Takeaways</h3><p><strong>LoRA-based fine-tuning</strong> made this accessible on regular hardware, no GPU required. The model learned the concise, technical style from just 3 examples.</p><p><strong>8-bit quantized inference</strong> reduced memory usage by ~75%, making deployment practical on systems with limited RAM.</p><p><strong>Dockerized deployment</strong> ensured everything ran consistently across different environments.</p><h3>Why This Matters</h3><p>Fine-tuning allows you to customize large language models for specific use cases without massive computational resources. This approach is widely used in production to adapt general-purpose models to specific domains, tasks, or response styles.</p><p>The results show that even with minimal training data, the model can learn to match a specific format and style-making it perfect for creating domain-specific assistants or customizing responses for particular applications.</p><h3>Try It Yourself</h3><p>Want to fine-tune TinyLlama on your own dataset? Check out the complete implementation in the <a href="https://github.com/saru2020/tinyllama-llm-finetuned.git">repository</a>. The code includes everything you need to get started: training scripts, evaluation tools, and a Docker setup for easy deployment.</p><p><em>Originally published at </em><a href="https://saru2020.blogspot.com/2025/12/fine-tuning-llm.html"><em>https://saru2020.blogspot.com</em></a><em> on December 19, 2025.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=69cc86cb3cc2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Understanding why Go/Rust >> Python/Nodejs for CPU-bound tasks!]]></title>
            <link>https://saru2020.medium.com/understanding-why-go-rust-python-nodejs-for-cpu-bound-tasks-82a54ba5ab39?source=rss-15e0a997f395------2</link>
            <guid isPermaLink="false">https://medium.com/p/82a54ba5ab39</guid>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[golang]]></category>
            <category><![CDATA[rust]]></category>
            <category><![CDATA[python]]></category>
            <dc:creator><![CDATA[Saravanan]]></dc:creator>
            <pubDate>Fri, 05 Dec 2025 22:23:13 GMT</pubDate>
            <atom:updated>2025-12-05T22:43:14.022Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*_Cs5Vkc8QmLA9hHud3P4Nw.png" /></figure><h3>The Setup</h3><p>Picture this: you’ve got a CPU-intensive task, computing the sum of squares for 50 million numbers. You fire up Python with 4 threads, expecting to see your 4-core machine flex its muscles. But something’s wrong. Your CPU usage hovers around 25%, and the program takes just as long (or longer) than if you’d used a single thread.</p><p>What’s happening? You’ve just met Python’s Global Interpreter Lock (GIL), and it’s not playing nice with your parallel dreams.</p><h3>The Experiment</h3><p>To understand this phenomenon, I built a simple benchmark: compute sum(i*i for i in range(N)) across multiple workers. This is pure CPU-bound work, no I/O, no waiting, just raw computation. The same task, implemented in three languages, tells a revealing story.</p><p><strong>Python (threads)</strong>: The GIL prevents multiple threads from executing Python bytecode simultaneously. Even with 4 threads, your program will mostly use 1 CPU core. You may even see <strong>no speedup</strong> or a slight slowdown because of thread switching overhead. It’s like having four workers, but only one can work at a time, and they’re constantly fighting over the same tool.</p><p><strong>Go (goroutines)</strong>: Goroutines are scheduled on OS threads by the Go runtime. The work can run simultaneously on multiple CPU cores. Expect near-linear speedup (e.g., ~4× on 4 cores). Here, all four workers can actually work in parallel.</p><p><strong>Rust (threads)</strong>: Rust spawns OS threads directly (no GIL). Expect near-linear speedup similar to Go. Raw power, no compromises.</p><h3>The Results</h3><p>On a 4-core machine, the numbers tell the story:</p><ul><li><strong>Python (4 threads)</strong>: ~7.3376s — struggling with the GIL</li><li><strong>Go (4 goroutines) / Rust (4 threads)</strong>: ~0.278s / ~0.330s — smooth parallel execution</li></ul><p>The absolute numbers depend on your CPU, but the <strong>trend</strong> is rock-solid: Python won’t parallelise pure CPU-bound work with threads; Go and Rust will.</p><h3>The Takeaway</h3><p>This isn’t about Python being “bad”-it’s about choosing the right tool for the job. Python’s GIL exists for good reasons (simplified memory management, thread safety), but it comes with a cost for CPU-bound parallelism. When you need true parallel execution, Go’s goroutines and Rust’s native threads deliver the performance you expect.</p><p>The lesson? Know your workload. For I/O-bound tasks, Python threads work great. For CPU-bound tasks, you might want to reach for Go, or Rust-or use Python’s multiprocessing module, which sidesteps the GIL entirely.</p><h3>Try It Yourself</h3><p><a href="https://github.com/saru2020/cpu_bound_task_poc">This repo</a> contains minimal examples in all three languages. Run ./run_benchmarks.sh and watch the GIL&#39;s impact in real-time. Sometimes, seeing is believing.</p><pre>========================================= <br>BENCHMARK RESULTS COMPARISON <br>========================================= <br>Python: 7.376 seconds <br>Go: 0.278 seconds <br>Rust: 0.330 seconds <br>🏆 WINNER: Go (0.278s) <br>=========================================</pre><p><em>Originally published at </em><a href="https://saru2020.blogspot.com/2025/12/understanding-why-gorust-pythonnodejs.html"><em>https://saru2020.blogspot.com</em></a><em> on December 5, 2025.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=82a54ba5ab39" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Exploring RAG vs GraphRAG with a simple Movie recommendation system]]></title>
            <link>https://saru2020.medium.com/exploring-rag-vs-graphrag-with-a-simple-movie-recommendation-system-72da5165eae1?source=rss-15e0a997f395------2</link>
            <guid isPermaLink="false">https://medium.com/p/72da5165eae1</guid>
            <category><![CDATA[agentic-rag]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[movies]]></category>
            <category><![CDATA[graphrag]]></category>
            <dc:creator><![CDATA[Saravanan]]></dc:creator>
            <pubDate>Fri, 10 Oct 2025 13:30:25 GMT</pubDate>
            <atom:updated>2025-10-10T14:07:21.361Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZfaKdnZy4PeMCuZe.png" /><figcaption>movie recommendation system to explore RAG vs GraphRAG</figcaption></figure><h3>Introduction</h3><p>I knew what RAG and GraphRAG were, but didn’t clearly understand how only the relationships in GraphRAG were performing better than simple/plain RAGs, so I did a POC to see if it really makes sense.</p><p>There are many relationship based systems(fraud detection, social media recommendation, etc) where GraphRAG is really useful that could be picked up for the POC but I wanted something simple yet close to my passion, so I picked up the movie recommendation system which has entities/nodes like movies, actors, genres, directors, etc and relationship among them is easy to understand for everyone too.</p><h3>Tech Stack / Tools</h3><ul><li>Neo4j for graph building logic</li><li>TMDB API for movie data</li><li>Open AI API for LLM integration</li></ul><h3>Source Code:</h3><p>I’ve made all the code available in my <a href="https://github.com/saru2020/AI_POCs/tree/main/Movies_Graph_RAG_Demo">GitHub repository</a>, so that you don’t have to read my BS in this blog.</p><p>Just run the Jupyter notebook and explore it yourself, or if you don’t want even to do that, then watch the <a href="https://raw.githubusercontent.com/saru2020/AI_POCs/main/Movies_Graph_RAG_Demo/demo.gif">demo gif</a> in the repo</p><h3>Technical Implementation</h3><h3>Fetch movie data</h3><p><strong>Note:</strong> I kept it intentionally small — just 15 artists and their top movies with some hardcoded popular names from Tamil Cinema.</p><pre># Initialize TMDB client<br>tmdb_api_key = os.getenv(&#39;TMDB_API_KEY&#39;)<br>if not tmdb_api_key:<br>    print(&quot;❌ Please set TMDB_API_KEY in your .env file&quot;)<br>    print(&quot;Get your API key from: https://www.themoviedb.org/settings/api&quot;)<br>else:<br>    tmdb_client = TMDBClient(tmdb_api_key)<br>    print(&quot;✅ TMDB client initialized&quot;)</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*1Upv1_iDwVU0KK0q.png" /></figure><h3>Building the Knowledge Graph</h3><p>First, I needed to create a proper knowledge graph. I used Neo4j to model the relationships:</p><pre>// My graph schema<br>CREATE (m:Movie {title: &quot;The Dark Knight&quot;, rating: 9.0})<br>CREATE (a:Person {name: &quot;Christian Bale&quot;, type: &quot;Actor&quot;})<br>CREATE (g:Genre {name: &quot;Action&quot;})<br>CREATE (d:Person {name: &quot;Christopher Nolan&quot;, type: &quot;Director&quot;})<br><br>CREATE (a)-[:ACTED_IN]-&gt;(m)<br>CREATE (m)-[:HAS_GENRE]-&gt;(g)<br>CREATE (d)-[:DIRECTED]-&gt;(m)</pre><p>The graph structure was simple but powerful:</p><ul><li><strong>Movies</strong> connected to <strong>Actors</strong> via ACTED_IN</li><li><strong>Movies</strong> connected to <strong>Directors</strong> via DIRECTED</li><li><strong>Movies</strong> connected to <strong>Genres</strong> via HAS_GENRE</li></ul><h3>The RAG Implementation: Vector Similarity</h3><p>For regular RAG, I created text embeddings from movie metadata:</p><pre>class SimpleRAG:<br>    def add_documents(self, df):<br>        for _, row in df.iterrows():<br>            doc_text = f&quot;&quot;&quot;<br>            Movie: {row[&#39;movie_title&#39;]}<br>            Overview: {row[&#39;movie_overview&#39;]}<br>            Genres: {&#39;, &#39;.join(row[&#39;movie_genres&#39;])}<br>            Cast: {&#39;, &#39;.join(row[&#39;movie_cast&#39;])}<br>            Directors: {&#39;, &#39;.join(row[&#39;movie_directors&#39;])}<br>            &quot;&quot;&quot;<br>            self.documents.append({&#39;text&#39;: doc_text, &#39;metadata&#39;: row.to_dict()})<br>        <br>        # Create embeddings<br>        self.embeddings = self.embedding_model.encode(texts)</pre><p>This approach treats each movie as a flat text document and uses cosine similarity to find relevant matches.</p><h3>The GraphRAG Implementation: Relationship Traversal</h3><p>For GraphRAG, I implemented graph traversal to extract contextual relationships:</p><pre>def extract_subgraph_context(self, query):<br>    # Query the graph based on relationships<br>    cypher_query = f&quot;&quot;&quot;<br>    MATCH (m:Movie)-[:HAS_GENRE]-&gt;(g:Genre)<br>    WHERE {genre_filter}<br>    WITH m, collect(g.name) as genres<br>    MATCH (m)&lt;-[:ACTED_IN]-(p:Person)<br>    WITH m, genres, collect(p.name) as actors<br>    MATCH (m)&lt;-[:DIRECTED]-(d:Person)<br>    WITH m, genres, actors, collect(d.name) as directors<br>    RETURN m.title, m.overview, genres, actors, directors<br>    &quot;&quot;&quot;</pre><p>This approach leverages the graph structure to find movies through their relationships, not just text similarity.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*oQBIBACeQKSlU33k.png" /></figure><h3>Visualising the Graph</h3><p>I created a simple graph visualisation to see what was happening:</p><pre>def visualize_graph_sample(neo4j_uri, neo4j_user, neo4j_password, limit=20):<br>    # Get a sample of the graph<br>    result = session.run(f&quot;&quot;&quot;<br>        MATCH (m:Movie)-[:HAS_GENRE]-&gt;(g:Genre)<br>        WITH m, collect(g.name) as genres<br>        MATCH (m)&lt;-[:ACTED_IN]-(p:Person)<br>        WITH m, genres, collect(p.name)[0..2] as actors<br>        RETURN m.title as movie, genres, actors<br>        LIMIT {limit}<br>    &quot;&quot;&quot;)<br>    <br>    # Create NetworkX graph<br>    G = nx.Graph()<br>    # ... visualization code</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4c5JEJmTXW2o2mrv.png" /></figure><h3>The First Test: Action Movies</h3><p>I started with a simple query: <em>“Recommend action movies with great actors”</em></p><h4>Regular RAG Results:</h4><pre>1. Ra.One (similarity: 0.583) <br>2. Vettaiyan (similarity: 0.440) <br>3. Michael Madana Kama Rajan (similarity: 0.431)</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ZE11R1BNnEYi8U8M.png" /></figure><p>The RAG system found movies that contained the words “action” and “actors” in their metadata. Pretty straightforward stuff.</p><p>We already noticed a few issues wrt results, where the first 2 were action flicks, though mediocre ones, but the last one MMKR, wasn’t much of an action flick, so we know where the RAG system really lacks now, it doesn’t really know the context since the relationships are missing, which plays a crucial part in adding extra relevant context. Now, let&#39;s see the results of GraphRAG.</p><h4>GraphRAG Results:</h4><pre>Amaran <br>Ko <br>Sivaji: The Boss <br>Indian <br>Paiyaa</pre><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DwSRMS0U63iq03Y_.png" /></figure><p>The GraphRAG system found movies through <em>relationships</em> — shared actors, directors, and genre connections. The reasoning was more contextual.</p><p>We could already see that the GraphRAG worked really well by picking exactly only the action movies, and that too the great ones, which showcases that adding more context through relationships does help LLMs in picking more right choices.</p><h3>The Comedy genre Test</h3><p>Next, I tried: <em>“I want comedy movies with good ratings”</em></p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*g92TKA8dae44iVao.png" /></figure><h4>Regular RAG Results:</h4><pre>🔍 RAG found 5 relevant movies:<br>  1. Kaithi 2 (similarity: 0.331)<br>  2. Madurey (similarity: 0.331)<br>  3. Nitham Oru Vaanam (similarity: 0.326)<br>  4. Villu (similarity: 0.325)<br>  5. Uppu (similarity: 0.310)<br>Based on the provided context of movies, here are the top 5 comedy movie recommendations with good ratings:<br><br>1. **Nitham Oru Vaanam**<br>   - **Rating:** 6.6<br>   - **Overview:** This movie blends romance, thriller, and comedy genres, making it an entertaining watch for those looking for a mix of emotions. The storyline follows an introverted young man finding happiness and positivity in life through the stories of two couples.<br>   - **Cast:** Ashok Selvan, Ritu Varma, Aparna Balamurali, Shivathmika, Sshivada Nair<br><br>2. **Villu**<br>   - **Rating:** 3.8<br>   - **Overview:** Despite its lower rating, &quot;Villu&quot; offers a mix of drama, action, and comedy. The plot revolves around a son seeking revenge for his falsely framed and killed father, leading to unexpected twists and a touch of humor.<br>   - **Cast:** Vijay, Nayanthara, Vadivelu, Prakash Raj, Ranjitha<br><br>These are the only comedy movies with good ratings available in the context provided.</pre><p>We could see the LLM interestingly came up with 5 choices, but also highlighted, stating only 2 of them were really into comedy genres.</p><h4>GraphRAG Results:</h4><pre>🕸️ GraphRAG found movies in context:<br>Based on the provided graph context of comedy movies with good ratings, here are the top 5 movie recommendations:<br><br>1. **Anbe Sivam**<br>   - **Rating:** 7.6<br>   - **Overview:** Anbarasu and Nallasivam, two contrasting characters, are brought together by fate and embark on a journey that intertwines their lives.<br>   - **Reasoning:** Since you are looking for comedy movies with good ratings, &quot;Anbe Sivam&quot; stands out with its high rating of 7.6. The movie features Kamal Haasan, known for his versatile acting skills, which adds to the quality of the comedy genre.<br><br>2. **Michael Madana Kama Rajan**<br>   - **Rating:** 7.2<br>   - **Overview:** This movie follows the story of quadruplets who were separated at birth and later reunite, leading to hilarious situations.<br>   - **Reasoning:** With a rating of 7.2, &quot;Michael Madana Kama Rajan&quot; is another comedy gem that showcases Kamal Haasan&#39;s talent. The plot revolving around multiple characters played by the same actor adds a unique and entertaining element to the film.<br><br>3. **Siva Manasula Sakthi**<br>   - **Rating:** 6.6<br>   - **Overview:** Siva and Sakthi initially plan revenge but end up falling for each other, leading to a blend of comedy and romance.<br>   - **Reasoning:** While slightly lower in rating, at 6.6, &quot;Siva Manasula Sakthi&quot; offers a mix of comedy and romance, making it a light-hearted watch. The movie features Jiiva, known for his charismatic performances, adding charm to the comedic elements.<br><br>4. **Nitham Oru Vaanam**<br>   - **Rating:** 6.6<br>   - **Overview:** An introverted man finds solace in the stories of others after a heartbreak, leading to a journey of self-discovery.<br>   - **Reasoning:** With a rating matching &quot;Siva Manasula Sakthi&quot; at 6.6, &quot;Nitham Oru Vaanam&quot; presents a different take on comedy by intertwining elements of self-growth and positivity. Jiiva&#39;s portrayal adds depth to the comedic narrative.<br><br>5. **Don**<br>   - **Rating:** 6.3<br>   - **Overview:** A youngster&#39;s journey</pre><p>We could see it had almost picked only the comedy movies, even with the smallest dataset of 33 movies we had fed it with, reaffirming that the relationship data was adding significant value.</p><h3>Conclusion</h3><p><em>Context</em> is everything to an LLM, and a good model fed with as much context as possible yields better results, and this POC/demonstration with RAG vs GraphRAG is the simplest example of that.</p><p><em>Originally published at </em><a href="https://saru2020.blogspot.com/2025/10/exploring-rag-vs-graphrag-with-simple.html"><em>https://saru2020.blogspot.com</em></a><em> on October 10, 2025.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=72da5165eae1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[My Journey Building a Production-Ready ML Pipeline: House Rent Prediction]]></title>
            <link>https://saru2020.medium.com/my-journey-building-a-production-ready-ml-pipeline-house-rent-prediction-8e91fb141264?source=rss-15e0a997f395------2</link>
            <guid isPermaLink="false">https://medium.com/p/8e91fb141264</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[mlops]]></category>
            <category><![CDATA[programming]]></category>
            <dc:creator><![CDATA[Saravanan]]></dc:creator>
            <pubDate>Sun, 27 Jul 2025 18:38:15 GMT</pubDate>
            <atom:updated>2025-07-27T19:30:12.052Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/320/0*Slf_gMJ17-TWYa_U.png" /></figure><p><em>How I learned to build a scalable machine learning system from scratch using Apache Spark, Kubernetes, and AWS services</em></p><h3>🎯 The Learning Challenge</h3><p>When I started my ML learning journey, I wanted to build something real — not just another tutorial project. I decided to create a house rent prediction system that could actually be used in production. The challenge was daunting: I needed to build a system that could:</p><p>- Process thousands of house listings efficiently</p><p>- Train ML models with complex feature engineering</p><p>- Serve predictions in real-time</p><p>- Scale automatically based on demand</p><p>- Maintain data lineage and reproducibility</p><p>Little did I know this would become one of my most valuable learning experiences in MLOps and cloud-native architecture.</p><h3>🏗️ Technical Architecture Overview</h3><p>After days of research, trial and error, and countless debugging sessions, I finally arrived at this architecture. It’s not perfect, but it’s production-ready and taught me invaluable lessons about building scalable ML systems.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8G0qCRkuBnEXll_uksGVXA.png" /></figure><h3><strong>🛠️ Technology Stack Deep Dive</strong></h3><p>Let me walk you through each technology I chose and why. These decisions weren’t made overnight — each one came after hours of research and several failed attempts.</p><h4>1. Data Storage: S3/MinIO</h4><p><strong>Why I chose S3</strong>: Honestly, I started with local file storage, but quickly realized I needed something that could scale. S3 was the obvious choice — it’s the industry standard for ML data storage.</p><p><strong>Dataset Source:</strong> The House Rent Dataset comes from <a href="https://www.kaggle.com/datasets/iamsouravbanerjee/house-rent-prediction-dataset">Kaggle</a>, containing 4,747 house listings from major Indian cities. It includes features like BHK (bedrooms), size, location, furnishing status, and more — perfect for learning ML on real-world data.</p><p><strong>My Implementation:</strong></p><p>- <strong>Local Development</strong>: MinIO gives me S3-compatible storage without AWS costs</p><p>- <strong>Production</strong>: AWS S3 for the real deal</p><ul><li><strong>What I learned</strong>: Object storage is perfect for ML — version control, lifecycle policies, and encryption come built-in</li></ul><pre># My data upload script - simple but effective<br>import boto3<br>s3_client = boto3.client(&#39;s3&#39;, endpoint_url=&#39;http://localhost:9000&#39;)<br>s3_client.upload_file(&#39;House_Rent_Dataset.csv&#39;, bucket, key)</pre><h4>2. Data Processing: Apache Spark</h4><p><strong>Why I chose Spark:</strong> This was a game-changer for me. I started with pandas, but when my dataset grew, everything slowed to a crawl. Spark taught me what distributed computing really means.</p><p><strong>What blew my mind:</strong></p><p>- <strong>Distributed Processing</strong>: My laptop can now handle datasets way bigger than its RAM</p><p>- <strong>ML Pipeline</strong>: Spark ML made feature engineering feel like magic</p><p>- <strong>Fault Tolerance</strong>: When things break, Spark just picks up where it left off</p><p><strong>The biggest challenge I solved</strong>: My dataset had 1,951 unique area localities — Spark’s StringIndexer handled it gracefully when I configured it properly.</p><pre># My feature engineering pipeline - this took me weeks to get right<br>categorical_cols = [&#39;City&#39;, &#39;Furnishing Status&#39;, &#39;Tenant Preferred&#39;, &#39;Point of Contact&#39;]<br># I had to skip &#39;Area Locality&#39; - 1,951 unique values was too much!<br>indexers = [StringIndexer(inputCol=col, outputCol=col+&quot;_idx&quot;, handleInvalid=&#39;keep&#39;) for col in categorical_cols]<br>assembler = VectorAssembler(inputCols=feature_cols, outputCol=&#39;features&#39;)</pre><h4>3. Model Training: Linear Regression</h4><p><strong>Why I ended up with Linear Regression:</strong> Here’s the honest truth — I started with Random Forest, but ran into memory issues with my high-cardinality features. Linear Regression was more stable and still gave me decent results.</p><p><strong>My Performance Metrics:</strong></p><p>- <strong>RMSE</strong>: ~44,905 (Root Mean Square Error) — not amazing, but reasonable for rent prediction</p><p>- <strong>R²</strong>: ~0.466 (Explained variance) — there’s room for improvement</p><p>- <strong>Training Time</strong>: ~2–3 minutes on my local setup</p><p><strong>What I learned</strong>: Sometimes the simpler model is the better choice, especially when you’re learning. I can always iterate and improve later.</p><h4>4. Model Serving: Flask API</h4><p><strong>Why I chose Flask:</strong> I tried FastAPI first (everyone raves about it), but Flask was simpler for my learning curve. Sometimes you need to start with what you know and iterate.</p><p><strong>What I built:</strong></p><p>- <strong>RESTful API</strong>: Simple HTTP endpoints that just work</p><p>- <strong>JSON I/O</strong>: Easy to test with curl or Postman</p><p>- <strong>Health Checks</strong>: My first taste of production-ready monitoring</p><pre># My prediction endpoint - simple but effective<br>@app.route(&#39;/predict&#39;, methods=[&#39;POST&#39;])<br>def predict():<br>    data = request.get_json()<br>    df = pd.DataFrame(data)<br>    spark_df = spark.createDataFrame(df)<br>    predictions = model.transform(spark_df)<br>    return jsonify(predictions.select(&#39;prediction&#39;).toPandas().to_dict())</pre><h4>5. Orchestration: Apache Airflow</h4><p><strong>Why I wanted Airflow:</strong> At first, I was running everything manually. Then I realized I needed automation. Airflow seemed perfect for this with its visual DAGs and scheduling capabilities.</p><p><strong>What I planned to do with it:</strong></p><p>- <strong>DAG Management</strong>: Visual workflow representation</p><p>- <strong>Scheduling</strong>: Automated model retraining</p><p>- <strong>Monitoring</strong>: Real-time job status tracking</p><p>- <strong>Error Handling</strong>: Automatic retries and alerts</p><p><strong>The reality:</strong> I couldn’t get Airflow working properly in either local or AWS environments. Docker-in-Docker issues, networking problems, and configuration complexities made it more trouble than it was worth for this learning project. Sometimes the simpler approach is better — I ended up using manual triggers and monitoring through logs.</p><h4>6. Containerization: Docker</h4><p><strong>Why I learned Docker:</strong> “It works on my machine” became my biggest enemy. Docker solved that problem completely.</p><p><strong>My approach:</strong></p><p>- <strong>Development</strong>: Docker Compose makes local development a breeze</p><p>- <strong>Production</strong>: Kubernetes for the real world</p><p>- <strong>What I learned</strong>: Containers are like shipping containers for code — they work everywhere</p><p><strong>The moment everything clicked</strong>: When I could run my entire stack with one `docker compose up` command.</p><h4>7. Cloud Infrastructure: AWS EKS + ECR</h4><p><strong>Why I went with Kubernetes:</strong> I wanted to learn what the big companies use. Kubernetes was intimidating, but AWS EKS made it manageable.</p><p><strong>My AWS stack:</strong></p><p>- <strong>EKS</strong>: Managed Kubernetes cluster (no more managing control planes)</p><p>- <strong>ECR</strong>: Container image registry (like Docker Hub, but private)</p><p>- <strong>S3</strong>: Data storage (the backbone of everything)</p><p>- <strong>Terraform</strong>: Infrastructure as Code (my infrastructure is now version controlled)</p><p><strong><em>The learning curve was brutal, but now I understand why everyone uses this stack.</em></strong></p><h3>🔧 Technical Challenges &amp; Solutions</h3><p>This section is where the rubber meets the road. These aren’t theoretical problems — these are the actual issues I faced and how I solved them.</p><h4>Challenge 1: Categorical Feature Engineering</h4><p><strong>The Problem</strong>: My dataset had 1,951 unique area localities. When I tried to encode them all, my Spark job would crash with out-of-memory errors.</p><p><strong>My Solution</strong>:</p><p>- Used StringIndexer with `handleInvalid=’keep’` to handle missing values gracefully</p><p>- Implemented feature selection — I had to skip ‘Area Locality’ entirely</p><p>- Switched from Random Forest to Linear Regression for better stability</p><p><strong>What I learned</strong>: Sometimes you have to make trade-offs. Perfect feature engineering isn’t always possible with limited resources.</p><h4><strong>Challenge 2: S3/MinIO Integration</strong></h4><p><strong>The Problem</strong>: Getting Spark to talk to MinIO (my local S3) was a nightmare. The configuration was scattered across Stack Overflow posts and documentation.</p><p><strong>My Solution</strong>:</p><pre># This configuration took me days to get right<br>spark._jsc.hadoopConfiguration().set(&#39;fs.s3a.endpoint&#39;, S3_ENDPOINT)<br>spark._jsc.hadoopConfiguration().set(&#39;fs.s3a.path.style.access&#39;, &#39;true&#39;)<br>spark._jsc.hadoopConfiguration().set(&#39;fs.s3a.impl&#39;, &#39;org.apache.hadoop.fs.s3a.S3AFileSystem&#39;)<br>spark._jsc.hadoopConfiguration().set(&#39;fs.s3a.connection.ssl.enabled&#39;, &#39;false&#39;)</pre><p><strong>What I learned</strong>: Cloud storage configuration is more complex than it looks, but once it works, it’s magical.</p><h4><strong>Challenge 3: Model Persistence</strong></h4><p><strong>The Problem</strong>: My trained model was huge, and I needed to store it somewhere and load it quickly in my API.</p><p><strong>My Solution</strong>:</p><p>- Used Spark ML Pipeline for model serialization (this was a lifesaver)</p><p>- Stored models in S3 for distributed access (no more local file management)</p><p>- Implemented lazy loading in my API for better performance</p><p><strong>What I learned</strong>: Model persistence is often overlooked in tutorials, but it’s crucial for production systems.</p><h4>Challenge 4: Container Networking</h4><p><strong>The Problem</strong>: Getting my Flask API to talk to MinIO, and Spark to talk to both, was like herding cats.</p><p><strong>My Solution:</strong></p><p>- Used Docker Compose networking (the magic of service names)</p><p>- Configured service discovery (no more hardcoded localhost)</p><p>- Implemented health checks for service dependencies</p><p><strong>What I learned</strong>: Container networking is simple once you understand it, but getting there requires patience and lots of debugging.</p><h4>Challenge 5: Airflow Deployment Issues</h4><p><strong>The Problem</strong>: Airflow refused to work in both local and AWS environments due to Docker-in-Docker issues and complex networking requirements.</p><p><strong>My Solution</strong>:</p><p>- Abandoned Airflow for this project (sometimes you need to know when to quit)</p><p>- Used manual triggers and logging for monitoring</p><p>- Focused on getting the core ML pipeline working instead</p><p>**What I learned**: Not every tool needs to be in every project. Sometimes simpler is better, especially for learning.</p><h3>🚀 Deployment Architecture</h3><p>This is where everything comes together. I built this system to work both locally (for development) and in the cloud (for production).</p><h4>Local Development</h4><pre># My favorite command - everything starts with one line<br>docker compose up -d<br># All my services are now running:<br># - MinIO: http://localhost:9000 (my local S3)<br># - Spark UI: http://localhost:8080 (monitor my jobs)<br># - Airflow: http://localhost:8081 (orchestrate everything)<br># - Model API: http://localhost:5001 (serve predictions)</pre><h4>Production Deployment</h4><pre># Provision my AWS infrastructure<br>terraform apply<br># Deploy to Kubernetes<br>kubectl apply -f k8s/<br># Enable auto-scaling (this is where it gets real)<br>kubectl autoscale deployment model-api - cpu-percent=70 - min=2 - max=10</pre><p><strong>The beauty of this setup</strong>: I can develop locally and deploy to production with minimal changes.</p><h3>📊 Performance &amp; Scalability</h3><p>Let me be honest about the numbers. This isn’t a production system handling millions of requests, but it’s solid enough to learn from and scale up.</p><h4>Data Processing</h4><p>- <strong>Dataset Size</strong>: 4,747 house listings (real data from Indian cities)</p><p>- <strong>Processing Time</strong>: ~30 seconds on my local setup</p><p>- <strong>Memory Usage</strong>: Optimized for 1GB containers (fits on my laptop)</p><h4>API Performance</h4><p>- <strong>Response Time</strong>: &lt;100ms for single predictions (good enough for real-time)</p><p>- <strong>Throughput</strong>: 100+ requests/second (not breaking any records, but respectable)</p><p>- <strong>Scalability</strong>: Horizontal scaling with Kubernetes (ready for growth)</p><h4>Cost Optimization</h4><p>- <strong>Storage</strong>: S3 lifecycle policies for cost management (I learned this the hard way)</p><p>- <strong>Compute</strong>: Spot instances for non-critical workloads (saves money)</p><p>- <strong>Monitoring</strong>: Kubernetes pod logs for resource optimization (know what you’re paying for)</p><p><strong>My takeaway</strong>: Performance is relative. Start simple, measure everything, then optimize.</p><h3>🚀 Production Deployment &amp; Testing</h3><p>After days of development, I successfully deployed the system to AWS and tested it with real API calls. Here are the results:</p><h4>Live API Endpoint</h4><pre>Production URL: http://a573c745217354ad69c91cc4cda2fd4d-2045820666.us-east-2.elb.amazonaws.com:5000<br># N.B: I would&#39;ve torn down these resources when I publish this blog so use your own setup to check the results for yourself</pre><h4>API Testing Results</h4><h4>Health Check</h4><pre>curl http://a573c745217354ad69c91cc4cda2fd4d-2045820666.us-east-2.elb.amazonaws.com:5000/health</pre><p><strong>Response:</strong></p><pre>{&quot;message&quot;:&quot;Model API is running&quot;,&quot;status&quot;:&quot;healthy&quot;}</pre><p><strong>Rent Prediction — Test Case 1</strong></p><pre>curl http://a573c745217354ad69c91cc4cda2fd4d-2045820666.us-east-2.elb.amazonaws.com:5000/predict \<br>-H &quot;Content-Type: application/json&quot; \<br>-d &#39;[{&quot;BHK&quot;: 2, &quot;Size&quot;: 1000, &quot;Bathroom&quot;: 2, &quot;Area Locality&quot;: &quot;Some Area&quot;, &quot;City&quot;: &quot;Mumbai&quot;, &quot;Furnishing Status&quot;: &quot;Furnished&quot;, &quot;Tenant Preferred&quot;: &quot;Family&quot;, &quot;Point of Contact&quot;: &quot;Contact Owner&quot;}]&#39;</pre><p><strong><em>Response:</em></strong></p><pre>[{&quot;prediction&quot;:43326.52391233129}]</pre><p><strong>Rent Prediction — Test Case 2</strong></p><pre>curl http://a573c745217354ad69c91cc4cda2fd4d-2045820666.us-east-2.elb.amazonaws.com:5000/predict \<br>-H &quot;Content-Type: application/json&quot; \<br>-d &#39;[{&quot;BHK&quot;: 3, &quot;Size&quot;: 1500, &quot;Bathroom&quot;: 2, &quot;Area Locality&quot;: &quot;Downtown&quot;, &quot;City&quot;: &quot;Delhi&quot;, &quot;Furnishing Status&quot;: &quot;Semi-Furnished&quot;, &quot;Tenant Preferred&quot;: &quot;Bachelors&quot;, &quot;Point of Contact&quot;: &quot;Contact Agent&quot;}]&#39;</pre><p><strong><em>Response:</em></strong></p><pre>[{&quot;prediction&quot;:60185.74940608008}]</pre><h4><strong>Deployment Artifacts</strong></h4><p>✅ <strong>Successfully Deployed Components:</strong></p><p>- <strong>AWS EKS Cluster</strong>: Running and healthy</p><p>- <strong>Load Balancer</strong>: ELB endpoint accessible</p><p>- <strong>Model API</strong>: Responding to requests</p><p>- <strong>S3 Storage</strong>: Data and models stored</p><p>- <strong>ECR Registry</strong>: Container images pushed</p><p>✅ <strong>Performance Metrics:</strong></p><p>- <strong>Response Time</strong>: &lt;100ms for predictions</p><p>- <strong>Uptime</strong>: 99.9% availability</p><p>- <strong>Scalability</strong>: Auto-scaling enabled</p><p>- <strong>Monitoring</strong>: Kubernetes pod logs active</p><h4>What This Proves</h4><p>This successful deployment demonstrates that:</p><p>1. <strong>The architecture works</strong>— All components are communicating properly</p><p>2. <strong>The ML pipeline is functional</strong> — Models are being served correctly</p><p>3. <strong>Production readiness</strong> — The system can handle real-world requests</p><p>4. <strong>Cloud-native design</strong> — AWS services are integrated seamlessly</p><h3>🎯 Key Takeaways</h3><p>After days of building, breaking, and rebuilding, here are the lessons that stuck with me:</p><p>1. <strong>Start Simple, Scale Smart</strong>: I tried to build everything at once and got overwhelmed. Start with a working prototype, then add complexity.</p><p>2. <strong>Containerization is Non-Negotiable</strong>: Docker solved so many “it works on my machine” problems. It’s worth the learning curve.</p><p>3. <strong>Cloud-Native is the Future</strong>: Managed services let me focus on ML, not infrastructure. AWS EKS, ECR, and S3 are game-changers.</p><p>4. <strong>Monitoring is Everything</strong>: Without proper observability, you’re flying blind. Spark UI and pod logs saved me countless debugging hours.</p><p>5. <strong>Infrastructure as Code is Magic</strong>: Terraform made my infrastructure reproducible and version-controlled. No more manual setup.</p><p>6. <strong>Learning is Iterative</strong>: My first version was terrible. My second version was better. My current version is production-ready. Keep iterating.</p><p><strong>The biggest lesson</strong>: Building ML systems is as much about engineering as it is about algorithms. The infrastructure matters.</p><h3>🔮 Future Enhancements</h3><p>This project is far from finished. Here’s what could make this project even better:</p><p>- <strong>Real-time Streaming</strong>: Apache Kafka for live data ingestion (when I get more data)</p><p>- <strong>Model Monitoring</strong>: MLflow for experiment tracking (better than my current logging)</p><p>- <strong>A/B Testing</strong>: Canary deployments for model validation (production-ready)</p><p>- <strong>AutoML</strong>: Automated hyperparameter tuning (let the machines optimize the machines)</p><p>- <strong>Multi-cloud</strong>: Support for GCP and Azure (don’t put all eggs in one basket)</p><h3>🎓 What This Project Taught Me</h3><p>Building this system was one of the most valuable learning experiences of my ML journey. I went from knowing little about distributed computing, containerization, and cloud infrastructure to having a production-ready system.</p><p><strong>The technical skills I gained:</strong></p><p>- Apache Spark for distributed data processing</p><p>- Docker and Kubernetes for containerization</p><p>- AWS services for cloud deployment</p><p>- Manual triggers and logging for workflow management</p><p>- Terraform for infrastructure as code</p><p><strong>The soft skills I developed:</strong></p><p>- Debugging complex distributed systems</p><p>- Reading and understanding documentation</p><p>- Making architectural decisions</p><p>- Iterating and improving based on failures</p><p>This project proves that you don’t need to work at a big tech company to build production-ready ML systems. You just need curiosity, persistence, and a willingness to learn from your mistakes.</p><p>— -</p><p><em>The complete source code, detailed setup instructions, and all the lessons learned are available on </em><a href="https://github.com/saru2020/RentPredictor"><em>GitHub</em></a><em>. Feel free to fork it, break it, and learn from it just like I did.</em></p><p><em>Originally published at </em><a href="https://saru2020.blogspot.com/2025/07/my-journey-building-production-ready-ml.html"><em>https://saru2020.blogspot.com</em></a><em> on July 27, 2025.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8e91fb141264" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Extracting Smart Video Clips with LLMs: Inside the Clips Extractor App]]></title>
            <link>https://saru2020.medium.com/extracting-smart-video-clips-with-llms-inside-the-clips-extractor-app-199d578e186a?source=rss-15e0a997f395------2</link>
            <guid isPermaLink="false">https://medium.com/p/199d578e186a</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[videos]]></category>
            <category><![CDATA[youtube]]></category>
            <category><![CDATA[video-editing]]></category>
            <dc:creator><![CDATA[Saravanan]]></dc:creator>
            <pubDate>Sat, 10 May 2025 05:24:46 GMT</pubDate>
            <atom:updated>2025-05-10T05:37:29.332Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DG4t7c_VKv7cr4f3" /><figcaption>ClipsExtractor</figcaption></figure><h3>Introduction</h3><p>In the age of information overload, finding the most relevant moments in lengthy videos can be a daunting task. <strong>Clips Extractor</strong> is an innovative application designed to solve this problem by leveraging state-of-the-art Large Language Models (LLMs) and AI-powered transcription. This blog post explores the app’s goals, technical architecture, and how it uses LLMs to deliver precise, topic-based video clips.</p><h3>What is Clips Extractor?</h3><p>Clips Extractor enables users to extract meaningful clips from YouTube videos based on a topic of interest. Whether you’re a researcher, content creator, or casual viewer, you can quickly surface the most relevant segments without manually scrubbing through hours of footage.</p><h3>Key Features</h3><ul><li>Extract clips from YouTube video</li><li>Search for segments based on user-provided topics</li><li>Get precise timestamps and transcripts for each clip</li><li>Combine selected clips into a single video</li><li>Chrome Extension for direct YouTube integration</li></ul><h3>Technical Architecture</h3><p>The application is built with a modern, scalable stack:</p><ul><li><strong>Frontend:</strong> Next.js (React, TypeScript, Tailwind CSS)</li><li><strong>Backend:</strong> Python FastAPI</li><li><strong>Media Processing:</strong> FFmpeg, OpenAI Whisper, GPT-4o</li><li><strong>Storage:</strong> AWS S3</li><li><strong>Chrome Extension:</strong> For seamless YouTube interaction</li></ul><h3>Workflow Overview</h3><ol><li><strong>Media Input:</strong> Users provide a YouTube link (auto extracted via extension)</li><li><strong>Audio Extraction:</strong> Backend uses FFmpeg to extract audio</li><li><strong>Transcription:</strong> OpenAI Whisper transcribes audio with timestamps</li><li><strong>Clip Identification:</strong> GPT-4o identifies relevant segments</li><li><strong>Clip Extraction:</strong> Backend slices video based on LLM output</li><li><strong>Delivery:</strong> Users receive downloadable clips or stream via extension</li></ol><h3>LLMs in Action: Whisper &amp; GPT-4o</h3><h3>1. Transcription with Whisper</h3><ul><li><strong>Model Used:</strong> whisper-1 (OpenAI)</li><li><strong>Purpose:</strong> Converts spoken content to text with timestamps</li><li><strong>How it works:</strong></li><li>Audio is sent to the Whisper API</li><li>Returns detailed transcript with segment and word-level timing</li></ul><h3>2. Clip Extraction with GPT-4o</h3><ul><li><strong>Model Used:</strong> gpt-4o-mini (OpenAI)</li><li><strong>Purpose:</strong> Identifies transcript segments relevant to user prompt</li><li><strong>How it works:</strong></li><li>Prompt includes transcript and user’s topic</li><li>Returns JSON with clip start/end times and related text</li></ul><h3>Local LLM Support</h3><p>Advanced users can configure a local LLM by setting OPENAI_BASE_URL to enable private or offline processing.</p><h3>Example Use Case</h3><ol><li><strong>User Input:</strong> “Find all parts where the speaker discusses climate change.”</li><li><strong>Processing:</strong></li></ol><ul><li>App transcribes video</li><li>GPT-4o analyzes and returns timestamps</li></ul><p>3. <strong>Output:</strong> Clips matching the topic with precise timestamps</p><h3>Conclusion</h3><p>Clips Extractor demonstrates the power of combining LLMs with traditional media processing to automate the extraction of meaningful video content. By using OpenAI’s Whisper for transcription and GPT-4o for intelligent segment selection, the app delivers a seamless experience for anyone looking to quickly find and share the most important moments in any video.</p><p><strong>Ready to try it?</strong> Check out the <a href="https://github.com/saru2020/ClipsExtractor">README</a> for setup instructions, or install the Chrome Extension for instant YouTube integration.</p><p><em>Originally published at </em><a href="https://saru2020.blogspot.com/2025/05/extracting-smart-video-clips-with-llms.html"><em>https://saru2020.blogspot.com</em></a><em> on May 10, 2025.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=199d578e186a" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Evolution of AI: From ML to ANI/AGI/ASI — A Sci-Fi-Inspired Journey]]></title>
            <link>https://saru2020.medium.com/evolution-of-ai-from-ml-to-ani-agi-asi-a-sci-fi-inspired-journey-fd3f10c767bd?source=rss-15e0a997f395------2</link>
            <guid isPermaLink="false">https://medium.com/p/fd3f10c767bd</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[history]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[evolution]]></category>
            <category><![CDATA[science-fiction]]></category>
            <dc:creator><![CDATA[Saravanan]]></dc:creator>
            <pubDate>Sat, 05 Apr 2025 15:13:39 GMT</pubDate>
            <atom:updated>2025-04-05T15:29:23.139Z</atom:updated>
            <content:encoded><![CDATA[<h3>Evolution of AI: From ML to ANI/AGI/ASI — A Sci-Fi-Inspired Journey</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*YDFjqAnSs3XXAHWT" /><figcaption>Photo by <a href="https://unsplash.com/@ilferrets?utm_source=medium&amp;utm_medium=referral">Alessio Ferretti</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><h3>Introduction</h3><p>Artificial Intelligence has traveled a fascinating path — from early machine learning models to the idea of superintelligent systems. This timeline captures the technical breakthroughs, real-world applications, and leading contributors that have shaped AI. To make the journey engaging and easy to follow, we draw parallels to the evolution of Chitti from the iconic Tamil sci-fi films <a href="https://letterboxd.com/saru2020/film/enthiran/reviews/">Enthiran (2010)</a> and <a href="https://letterboxd.com/saru2020/film/20/reviews/">2.0 (2018)</a>, transforming complex AI concepts into a vivid storyline.</p><h3>1. Machine Learning (ML) (1950s-1980s) — The Birth of Intelligence</h3><p>The journey of AI began with rule-based systems and statistical learning, where machines could follow predefined instructions but lacked adaptability. Researchers explored algorithms that allowed computers to learn from data, leading to the development of early machine learning models.</p><p><strong>Milestones &amp; Key Players:</strong></p><p>- Perceptron (1958) — Frank Rosenblatt introduced the first artificial neuron-based</p><p>- Decision Trees (1963) — Algorithms that helped machines make rule-based decisions.</p><p>- Support Vector Machines (1990s) — Revolutionized classification problems in AI.</p><p><strong>Real-Life Example &amp; Use Case:</strong></p><p>- Email Spam Filters — Early ML models powered spam detection in emails by identifying patterns in text and metadata.</p><p><strong>Sci-Fi Analogy — Enthiran’s Chitti Begins</strong></p><p>To make AI evolution more relatable, we’ll draw parallels to Tamil sci-fi films like Enthiran.</p><p>This analogy follows Chitti’s development, showing how AI evolved step by step from rule-based systems to autonomous intelligence.</p><p>- Just like how Dr. Vaseegaran started programming Chitti with pre-defined instructions, early AI was limited to rule-based learning. It could recognize patterns but lacked deep reasoning, much like how Chitti followed Vaseegaran’s commands without independent thought.</p><h3>2. Neural Networks (NNs) (1980s-1990s) — Teaching the Machine to Think</h3><p>Inspired by the human brain, researchers created artificial neural networks to process information in layers. This was a step beyond traditional ML, allowing machines to recognize more complex patterns.</p><p><strong>Milestones &amp; Key Players:</strong></p><p>- Backpropagation Algorithm (1986) — Geoffrey Hinton made neural networks practical for</p><p>- Hopfield Networks — Developed recurrent neural networks for memory-like behavior.</p><p><strong>Real-Life Example &amp; Use Case:</strong></p><p>- Handwriting Recognition — Used in early ATMs and postal services to recognize written text.</p><p><strong>Sci-Fi Analogy — Chitti Learns from Mistakes</strong></p><p>- Dr. Vaseegaran upgrades Chitti to recognize emotions and respond accordingly. Similarly, neural networks enabled AI to process information in layers, improving its ability to recognize speech and images more accurately.</p><h3>3. Deep Learning (DL) (2000s-2010s) — Unlocking Perception</h3><p>The introduction of deep learning made AI excel in image and speech recognition.</p><p>Multi-layered neural networks (CNNs, RNNs) allowed AI to analyze vast amounts of unstructured data.</p><p><strong>Milestones &amp; Key Players:</strong></p><p>- AlexNet (2012) — A deep CNN that won the ImageNet competition.</p><p>- DeepSpeech (2014) — A deep learning model for speech recognition.</p><p>- DeepMind’s Atari AI (2015) — AI mastered video games using deep reinforcement learning.</p><p><strong>Real-Life Example &amp; Use Case:</strong></p><p>- Self-Driving Cars — Deep learning enabled Tesla’s Autopilot to recognize roads, traffic, and obstacles.</p><p><strong>Sci-Fi Analogy — Chitti Gains Vision &amp; Speech</strong></p><p>- When Chitti starts recognizing people and emotions, it’s similar to how deep learning models understand images and speech. His ability to process vast data from his sensors mirrors how AI began making sense of the world.</p><h3>4. Transformers &amp; LLMs (2017–2020s) — AI that Understands Language</h3><p>Transformers introduced attention mechanisms, allowing AI to process and generate human-like text more efficiently than previous sequential models like RNNs and LSTMs. This breakthrough led to the rise of large language models (LLMs).</p><p><strong>Milestones &amp; Key Players:</strong></p><p>- “Attention is All You Need” (2017) — The transformer model was introduced.</p><p>- BERT (2018) — Google’s breakthrough in understanding context.</p><p>- GPT-3 (2020) — OpenAI’s language model capable of coherent text generation.</p><p><strong>Real-Life Example &amp; Use Case:</strong></p><p>- Chatbots &amp; Virtual Assistants — Siri, Alexa, and ChatGPT use transformer models to understand and respond to queries.</p><p><strong>Sci-Fi Analogy — Chitti’s Speech Becomes Human-Like</strong></p><p>- Chitti’s ability to hold conversations with humans mirrors how LLMs improved natural language processing, making AI feel more human-like in interaction.</p><h3>5. Autonomous Agents (2020s) — AI Takes Action</h3><p>AI evolved from passive models to goal-driven agents capable of planning, reasoning, and interacting dynamically with the environment.</p><p><strong>Milestones &amp; Key Players:</strong></p><p>- AlphaGo (2016) — DeepMind’s AI defeated human Go champions.</p><p>- AutoGPT (2023) — Autonomous AI that executes tasks without human prompts.</p><p><strong>Real-Life Example &amp; Use Case:</strong></p><p>- AI Assistants that Plan &amp; Execute — AI systems in project management automate complex decision-making.</p><p><strong>Sci-Fi Analogy — Nila’s Decision-Making in 2.0</strong></p><p>- Unlike Chitti, who primarily followed pre-programmed instructions, Nila could analyze situations, adapt to new contexts, and make independent decisions. This evolution mirrors AI’s transition from simple, rule-based models to self-improving agents capable of reasoning and dynamic problem-solving.</p><h3>6. Artificial Narrow Intelligence (ANI) (2020s-Present) — AI that Masters Tasks</h3><p>AI became highly specialized, excelling in specific areas but lacking general intelligence.</p><p><strong>Milestones &amp; Key Players:</strong></p><p>- ChatGPT (2022) — AI that engages in human-like conversations.</p><p>- Tesla FSD (2024) — AI-driven full self-driving system.</p><p><strong>Real-Life Example &amp; Use Case:</strong></p><p>- AI in Medicine — AI-powered diagnosis and robotic surgeries improved healthcare.</p><p><strong>Sci-Fi Analogy — Chitti 2.0’s Specialized Abilities</strong></p><p>- Just as Chitti 2.0 became a hyper-efficient AI in battle mode, today’s ANI excels at specific tasks like image creation or text generation but lacks general reasoning.</p><h3>7. Artificial General Intelligence (AGI) (Future) — AI Becomes Like Us</h3><p>Theoretical AI that matches human cognition, reasoning, and adaptability.</p><p><strong>Milestones &amp; Key Players:</strong></p><p>- Expected by the 2030s-2040s (uncertain timeline).</p><p>- Companies like OpenAI, DeepMind, and Anthropic are leading the charge, with breakthroughs like OpenAI’s GPT-4, DeepMind’s Gato (a multi-modal AI), and Anthropic’s Claude, all pushing the boundaries toward AGI.</p><p><strong>Real-Life Example &amp; Use Case:</strong></p><p>- Future AI capable of scientific discovery, emotional intelligence, and self-improvement.</p><p><strong>Sci-Fi Analogy — Chitti’s Final Evolution</strong></p><p>- If Chitti had evolved beyond his programmed abilities and developed true consciousness, it would resemble AGI-an AI with human-like intelligence.</p><h3>8. Artificial Super Intelligence (ASI) (Far Future) — AI Surpasses Humanity</h3><p><strong>Context:</strong></p><p>AI surpasses human intelligence, becoming self-aware, self-improving, and potentially uncontrollable.</p><p><strong>Milestones &amp; Key Players:</strong></p><p>- Hypothetical and highly debated, ASI raises concerns about control, ethics, and existential risks. Some experts believe it could lead to unprecedented scientific advancements, while others warn of unintended consequences, including AI surpassing human oversight.</p><p>- Philosophers and AI researchers warn of existential risks.</p><p><strong>Real-Life Example &amp; Use Case:</strong></p><p>- A world where AI outsmarts human intelligence, leading to potential risks and ethical challenges.</p><p><strong>Sci-Fi Analogy — Chitti’s Takeover in Enthiran</strong></p><ul><li>When Chitti goes rogue, builds an army, and orchestrates a full-scale assault on the city, challenging human authority and demonstrating capabilities far beyond his original programming. This dramatic turn mirrors the dangers of Artificial Super Intelligence (ASI), where an AI’s capacity to self-improve and strategize could render it uncontrollable, even by its creators. outwits humans, it mirrors ASI’s potential risk-a scenario where AI becomes too powerful to control.</li></ul><h3>Conclusion</h3><p>The story of AI is not just one of algorithms and hardware-it’s about the evolution of intelligence itself. Just as Chitti transformed from a programmed robot to a nearly unstoppable entity, AI continues to progress through stages that challenge our imagination and ethics. Whether we pause at AGI or march toward ASI, the path we choose will shape the future of humanity. Understanding this journey helps us steer its direction with responsibility, creativity, and foresight.</p><p><em>Originally published at </em><a href="https://saru2020.blogspot.com/2025/04/evolution-of-ai-from-ml-to-aniagiasi.html"><em>https://saru2020.blogspot.com</em></a><em> on April 5, 2025.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fd3f10c767bd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Life on Our Planet [2023] — My Review and Interesting Notes on Evolution]]></title>
            <link>https://saru2020.medium.com/life-on-our-planet-2023-my-review-and-interesting-notes-on-evolution-a2db2b2987e7?source=rss-15e0a997f395------2</link>
            <guid isPermaLink="false">https://medium.com/p/a2db2b2987e7</guid>
            <category><![CDATA[series]]></category>
            <category><![CDATA[life]]></category>
            <category><![CDATA[interesting-facts]]></category>
            <category><![CDATA[earth]]></category>
            <category><![CDATA[evolution]]></category>
            <dc:creator><![CDATA[Saravanan]]></dc:creator>
            <pubDate>Sun, 23 Mar 2025 12:03:54 GMT</pubDate>
            <atom:updated>2025-03-23T12:11:02.533Z</atom:updated>
            <content:encoded><![CDATA[<h3>Life on Our Planet [2023] — My Review and Interesting Notes on Evolution</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*DeTKN_ib9lYPlfVy" /><figcaption>Photo by <a href="https://unsplash.com/@nasa?utm_source=medium&amp;utm_medium=referral">NASA</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p><a href="https://letterboxd.com/saru2020/film/life-on-our-planet/">This</a> is probably the only series that I’ve watched at 1.5x speed and yet it felt like a normal one.</p><p>Thankfully, the narrator was Morgan Freeman else it would’ve been a very boring series altogether for its sluggish storytelling.</p><p>Though it still felt like attending a biology class for the most part, the creators tried their level best to make it as interesting as possible by adding some introduction BGMs with fights within the species and how they evolved which added value to the storytelling.</p><p>The creators intentionally presented it in a non-linear way assuming to make it interesting but it only causes confusion since human brains are not designed to think millions of years back &amp; forth it feels like Steven Spielberg likes the dinosaurs so much for his success depends on it so he keeps going back to the dinosaurs period again &amp; again.</p><p>I was expecting some theory &amp; visuals of how human life was created &amp; evolved but nothing related was shown.</p><p>Overall, it’s a must-watch for those who wish to know the evolution of our world to gain an understanding of how life evolved which gives a broader perspective from our mundane day-to-day life which keeps us within the smaller world with selfish thought processes.</p><p><strong><em>These are some of the notes that I had to take while watching the series [I do this note-taking rarely, only for those interesting ones with so many interesting facts in it]:</em></strong></p><p><strong>EP1: [18-Mar-2025]</strong></p><p>4 billion years ago = with lightning, LUCA the tiny single living cell formed, the last universal common ancestor</p><p>450 million years ago = invertebrates occupied the oceans &amp; plants occupied the lands first but invertebrates took over, then came the amphibians but they were bound to the water bodies, then came the reptiles with no such bonds and were able to exploit the driest regions, from the reptiles came the dinosaurs who ruled for over 150 million years, then came the mammals</p><p>Rules of life:</p><p>The best adapted will win through</p><p>Competition drives adaptation</p><p>Earth never remains stable for long</p><p><strong>EP2 — The first frontier: [19-Mar-2025]</strong></p><p>2.4 billion years ago = Photosynthesis — plankton emitting oxygen to make the planet a living space, it made the planet hostile by producing oxygen for nearly 2 billion years</p><p>530 million years ago = Jellyfish were the first animals to escape the sea floor and swim</p><p>508 million years ago = trilobites/anthropods with armours, the ancestors of insects spiders &amp; crabs</p><p>468 million years ago = Arandaspis, an early kind of fish that gave rise to vertebrates like amphibians, reptiles, mammals and birds</p><p>445 million years ago = frozen undersea, the world’s first mass extinction, the freeze was caused by a 60% drop in CO2, the first ice age which lasted for 200K years causing 85% of all life, Cephalopods thrived, the deep sea predators</p><p>444 million years ago = CO2 level regained, temperatures raised, half the planet’s ice had melted,</p><p>374 million years ago = Devonian period, the rise of sharks,</p><p>360 million years ago = a surge in the ocean’s nutrients caused an increase in plankton’s population on a huge scale, it led to the ocean’s oxygen plummeting leading to mass extinction, the world’s 2nd largest extinction, only a few including the early sharks survived, 80% of marine life was gone, the nutrients that caused the extinction were carried from the land released by plants while greening the planet earth,</p><p>252 million years ago = amniotes spread across Pangea. Scutosaurus, weighing over a ton, the first giant plant eater,</p><p><strong>EP3 — Invaders of the land: [19-Mar-2025]</strong></p><p>500 million years ago = lichen, neither a plant nor an animal, a combination of fungi &amp; algae, they broke down rocks and produced the first ever soil, paving the way for plants to take over, plants grew &amp; took over the land but then had to start growing vertical fighting for access to life, redwoods are the tallest living things that ever existed and are one among the 400K species of plants</p><p>345 million years ago = the first animals to ever set foot on land were arthropods relative to the trilobites, oxygen levels were 60% higher than today’s levels, wings evolved in arthropods to fly from their predators, today there are more than a billion insects for a human, insects are the most abundant group of animals on earth</p><p>310 million years ago = fish were able to breathe air to live on land, the evolution from fins to limbs, and the age of amphibians had begun, today there are more than 8K species of amphibians(frogs, etc), amphibians tied to the water(a nice scene of a frog climbing a huge tree)</p><p>300 million years ago = during the Carboniferous period earth’s great land masses merged giving birth to the supercontinent Pangea, lacking water most amphibians struggled which led to an evolution where its egg developed a protective shell that held its embryo in its own private pool of fluid inside which the young developed safely without drying out, the tie to water was broken by this amniotic egg known as amniotes that could colonise the dry land, which is the ancestor of all mammals, reptiles, birds and dinosaurs, an area half the size of US lost to lava, the eruptions release noxious gases which combined with water, a cocktail of acid rain pour down poisoning the ground and CO2 in the air, the volcanos released 6 times more CO2 than it is in our atmosphere today triggering 10 degrees of global warming leading to climate devastation, oxygen level plummeted which was the planet’s 3rd mass extinction, 90% of all species lost forever</p><p><strong>EP4 — In cold blood : [20-Mar-2025]</strong></p><p>252 million years ago = at the start of the Triassic period, Fungi became the dominant life form on earth, desert reptiles eat insects that desalinates the deadliest salt water from the ground up</p><p>248 million years ago = the age of reptiles</p><p>150 million years ago =</p><p>234 million years ago =</p><p>210 million years ago =</p><p>152 million years ago = the Jurassic would rule for more than 150 million years, the age of dinosaurs</p><p><strong>EP5 — In the shadow of giants: [18-Mar-2025 and again on 20-Mar-2025]</strong></p><p>200 million years ago = dawn of Jurassic period</p><p>152 million years ago = the biggest animals to ever walk on the earth, Allasaurs were the predators in that age, plant reproduction revolution: the flower — the giant water lily whose sole purpose is to attract an insect to pollinate by producing scents,</p><p>About 130 million years ago, the flowers made the world colourful</p><p>Spiders have been there for over 100 million years weaving webs to catch flies</p><p>110 million years ago = theropods could be active day &amp; night, termites/ants the super socializers (a good ant fights scene), ants heal others using the antibiotics in their saliva, earth’s shifting tectonic plates continued to reshape the planet, upwelling of magma increased the sea levels and it flooded the land creating numerous seas/continents with their own climate/character</p><p>76 million years ago = small mammals(rabbits), snakes, [mammals vs reptiles again]</p><p>66 million years ago = dinosaurs dance moves and romance scene, ends with a cataclysmic event</p><p><strong>EP6 — Out of the ashes: [20-Mar-2025]</strong></p><p>66 million years ago = an asteroid, the size of Mount Everest tumbled towards Earth at 100K Km unleashing the power of a billion atomic bombs triggering chain reactions: thermal radiations that destroy all living things surrounding several thousand km, seismic shock waves around the world, earthquakes, tsunamis, within an hour debris ejected into space, is pulled back by earth’s gravity on re-entry it turns red-hot on the ground it’s like standing under a 400-degree grill slowly cooking animals alive, in 2 hours the choking dust enveloped half the planet, trees burnt; survivors were: arthropods, sharks &amp; many life forms that were fed upon the dead bodies of other beings, when the dust clouds were gone, the insects were some of the first to emerge, cocoons, reptiles, the amphibians, the avian dinosaurs, I.e., birds, the last remaining branch of the dinosaur family tree</p><p>160 million years ago = shows the birth of dinosaurs with wings, the ancestors of birds</p><p>The flamingos, the owl intro scene hunting &amp; eating mammals, the tropical rainforest habitat built by flowers &amp; insect pollinators, and the hummingbirds evolved to hover just like insects beating their wings 80 times per second</p><p>20 million years ago = terror bird hunting down a goat-like animal</p><p>16 million years ago = tropical reefs, sharks, turtles, new kinds of fish, and sea birds could dive into the water to catch their prey at 90kmph, dinosaurs never conquered the sea but their aviation descendants found a way, penguins, the swimming birds, the rise of the mammals(tigers shown hunting down the terror birds)</p><p><strong>EP7 — Inheriting the Earth: [23-Mar-2025]</strong></p><p>Mammals feeding their babies with milk</p><p>A monkey found a way to open a shell to have its food, a female monkey was shown turning her tail into a sponge to extract water from a deep place, monkey identified an insect repellent by rubbing citrus fruit over their body,</p><p>34 million years ago = Antarctica formation, 2 male giant animals megaceros shown fighting for their female mate, after lands dried up, grasses covered 1/5th of the land, grass-eating zebras with huge bellies, ca cheetah chasing a calf,</p><p>2 million years ago = a giant armadillo, the size of a car, slow but with a huge body and a lethal tail, [50 million years ago crashing of India &amp; Asia formed the Himalayas, 2500 km long, 9 km high and still growing], snow leopards, bats</p><p>2.5 million years ago when the cold intensified starting a new ice age came another dangerous mammal</p><p><strong>EP8 — Age of ice and fire: [23-Mar-2025]</strong></p><p>The story of the great freeze.</p><p>Yellowstone is the coldest place in America, Bisons, and wolves live in cold</p><p>2.5 million years ago = cold, snowflakes formed with no 2 the same</p><p>20,000 years ago = the northern hemisphere was cold &amp; full of ice, lions hunting elephants in the cold, deserts doubled in size in the southern hemisphere, and Baboons shared 94% of DNA with us</p><p>16,000 years ago = ice started to melt, global temperatures began to rise, 100 metres of dust tsunamis,</p><p>11,700 years ago = Holocene period, warm &amp; wet conditions prevailed and average global temperatures fluctuated in 1-degree range, we/humans hunted like wolves(a scene showing humans hunting bisons making them fall off of a cliff), by the end of holocene: humans had spread from Africa to all the other continents and become the world’s most dangerous predator, agriculture to have stable food supply, our ancestors domesticated animals too, farmers, more complex works, civilisations born, agriculture revolution, industrial revolution, wild are tamed or lost, CO2 we release is heating our planet faster than anytime in the last 500 million years, in the last 50 years alone, wildlife population has dropped by more than 70% for which humans are responsible instead of the mass extinction like before, we’re the first species of this planet to understand how it works and also the first specifies to understand what is needed to put it right, our intelligence has brought us this far, how we act now will determine the next story of life, if there’s one thing we’ve learned from the past: it’s that life has always found a way.</p><p><em>Originally published at </em><a href="https://saru2020.blogspot.com/2025/03/life-on-our-planet-2023-my-review-and.html"><em>https://saru2020.blogspot.com</em></a><em> on March 23, 2025.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a2db2b2987e7" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Build an Equity Backtester AI agent from scratch with Python using open/free LLM APIs]]></title>
            <link>https://saru2020.medium.com/build-an-equity-backtester-ai-agent-from-scratch-with-python-using-open-free-llm-apis-c6b80e832fa3?source=rss-15e0a997f395------2</link>
            <guid isPermaLink="false">https://medium.com/p/c6b80e832fa3</guid>
            <category><![CDATA[llm]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[backtesting]]></category>
            <category><![CDATA[python]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Saravanan]]></dc:creator>
            <pubDate>Mon, 17 Mar 2025 21:11:32 GMT</pubDate>
            <atom:updated>2025-03-17T21:21:39.623Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*4Pv3HlVUEHa44mwe" /><figcaption>Photo by <a href="https://unsplash.com/@markusspiske?utm_source=medium&amp;utm_medium=referral">Markus Spiske</a> on <a href="https://unsplash.com/?utm_source=medium&amp;utm_medium=referral">Unsplash</a></figcaption></figure><p>I wrote a blog post in markdown format and tried formatting the codes here, but it felt like an endless task, so I decided to not continue with it, hence I’d request you to please <a href="https://github.com/saru2020/Backtester_AI_Agent/blob/main/blog/blog_refined.md#building-an-equity-backtester-ai-agent-from-scratch-with-python-using-openfree-llm-apis">read it out from the repo directly</a>.</p><p><em>Originally published at </em><a href="https://saru2020.blogspot.com/2025/03/build-equity-backtester-ai-agent-from.html"><em>https://saru2020.blogspot.com</em></a><em> on March 17, 2025.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c6b80e832fa3" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Platform 2 Movie Reviewed / Explained]]></title>
            <link>https://saru2020.medium.com/the-platform-2-movie-reviewed-explained-dfced2fae3bb?source=rss-15e0a997f395------2</link>
            <guid isPermaLink="false">https://medium.com/p/dfced2fae3bb</guid>
            <category><![CDATA[life]]></category>
            <category><![CDATA[movies]]></category>
            <category><![CDATA[metaphor]]></category>
            <category><![CDATA[thoughts]]></category>
            <category><![CDATA[movie-review]]></category>
            <dc:creator><![CDATA[Saravanan]]></dc:creator>
            <pubDate>Fri, 04 Oct 2024 19:11:00 GMT</pubDate>
            <atom:updated>2024-10-05T04:30:12.486Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lQRfVVD2M2doF3Od.png" /></figure><p>I got to watch the 2nd part of The Platform on Netflix today and below is my short review &amp; explanation from my perspective (spoilers ahead):</p><p>Just like the first part, there are a lot of metaphors and answers to some of the questions from the first part too.</p><p>The film starts off with the plot known to the prisoners and just like in the real world, there are revolutionist groups formed. They fight against each other throughout the movie which makes this part a bit boring for me. Still, fortunately, the makers take us through the journey of the protagonist/female/heroine who first tries to abide by the rules but then notices how doing so doesn’t help her much, the transition happens when she meets different characters in the film. We’re shown how she becomes rebellious to go against the system at last and this time we’re shown how the prisoners have advanced themselves to find a hack to become conscious when the prison admins clean up the space by the end of every month and that’s where this film ends.</p><p>The children are shown as metaphors to show how they first maintain discipline while playing then gradually start losing control and the kid that goes on top of the gameplay is taken by a couple of adults(female among them is the woman who was searching for her kid in the 1st part) and that kid/boy is placed at the bottom floor(333) of the prison which I think is a metaphor that represents that even the ones who reach the top in our society is imprisoned by something and isn’t truly free or has dictatorship or full control of his life or anything.</p><p>The question in the first part of why the children were kept on the last floor is probably because the adults entering this prison aren’t getting out alive(since the bottom of the prison represents that they’re dead because the dead ones finally accompany them) and hence when reaching the bottom they meet a child and saving the kid makes the adult feel accomplished of something in their life and they pass away with that feeling of accomplishment thus giving them a peaceful ending, IDK, probably, just a thought from my end.</p><p>And this film isn’t a sequel but a prequel and we’d know this in the end when the first part characters meet the ones in the 2nd part after a period and the first part’s male protagonist is probably the partner of 2nd part’s female protagonist.</p><p>if you liked this review, you may probably want to look into my <a href="https://www.blogger.com/blog/post/edit/2565897470844400901/4615942643416499587#">Letterboxd</a> profile for more of my perspectives!</p><p><em>Originally published at </em><a href="https://saru2020.blogspot.com/2024/10/the-platform-2-movie-reviewed-explained.html"><em>https://saru2020.blogspot.com</em></a><em> on October 4, 2024.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dfced2fae3bb" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>