Amuthan Ganeshan

MCP - a open protocol for AI (LLM) to talk to other tools / systems / applications. Naive, but don’t judge me, I was judging AI capabilities with me being Naive   - So I asked Claude AI to create an MCP - An MCP the AI itself will use - Gave a website address to which MCP had to be created (with public APIs) - It Generated extensive, detailed code - I dwelled on it, fine tuned for my preference / requirements, it went ahead and did a bit more than I asked for. Website was analysed in minutes (amusing and scary) - I asked it to package it as executable package - I had configured file system MCP - So, Claude leveraged that and created a python project in filesystem and created Project with proper files and folder structure - Here comes Claude code (CC) - This python project of MCP server, was not in a running state - Asked CC to make the code runnable (of course I was in control (manual loop) of the changes it was making) - MCP server to run in local system (MacOs) needed some changes in trivial python code, CC made the changes - Interesting side part, during this whole process, it was also creating documentation, to install server, to run server and the MCP config for other tools - At times I had to say update the documentation - OOPS !! Code got changed, but what was changed ? - I asked CC to make this project version controlled. Oh my Devaa (God), it converted the project into a Git repo (local) - Created git related files, Auto committed all files. All I had to do was, say yes, go ahead - Well, we got side tracked, lets get back to MCP - So, now I asked CC to run the server (lazy to copy the command and run it in next tab terminal) - Obviously.. It had issues, and obviously it auto corrected the code - It wrote test cases, wrote test start script, tested it, committed the code and then - now it tells me, 'All fixed, Want to run?', Howdhu maga howdhu, run (yes dude yes, run) - MCP server started. I updated MCP configuration and restarted Claude Desktop - pssssttt….. curios? - :( Error: MCP server couldn’t be started - Again me being extremely (purposefully) lazy, told CC, you are being dumb, fix the startup error - I guess, I hurt its Ego. - It asked permissions for MCP server log file. whaaaat really ?, it wants to read log files now, wow. - It test ran with its test scripts, read the log, analysed like a nerd, updated the test files and code, reran until it was fixed and updated the documentation and also committed the code - What can I say, I was watching Christopher Nolan movie level events. - Finally, MCP worked, I tested few things, results were as expected.   It ain’t over, my architectural / Engineering hormones kicked in. How about containerisation ? How about productionising it ? How about Security & Authentication ? How about Performance & Scalability ? How about Operational Readiness ? How about AI Security, Gaurdrails ? Stay tuned, I will be back 🙂

14

From ShadowTraffic, into MySQL, into Kafka using Source Connector (& a SMT), filtered, packaged & sinked into REDIS using Kafka Sink Connector (& another SMT) Overview 1. We start by first generating records using Shadowtraffic. 2. Insert the records created into a MySQL datastore. 3. We then utilise Kafka Connect Source Connector with a custom SMT to consume the records from our MySQL datastore, filtering out specific records, after adding a custom Message Key (not based on a column) as we publish onto specified Apache Kafka topic. 4. From here we use a Kafka Connect Sink Connector, again with a custom SMT function to filter our records, package (specified columns) into a Key/Value JSON payload into a REDIS datastore. Jay Prag Alvin Munsamy Gareth Tennant Gopolang Lekoto Jack Bingham Jaco Hennig Jeremy Rivera, CPWA Saleha Saib Nabbie Zaahida Hassen 🐿️ Ben Gamble🧑🏾‍🦯 Jark Wu Alex Merced Mehul Batra Mfundo Vinjwa Thobekani Tk Ndlovu Mary-Ann Jika Lesetja Charles Magongwa Hennie Human Hennie Venter Dion Guy Michael Drogalis Michelle Padayachy Jacques Fouché Odette Smit Confluent apache kafka Kameshwar Sorda Chris Norton Basie Kgaphola Claudine Els Peter Dixon Peter Pugh-Jones Kim Ormiston Felicia Dickerson Ricardo Ferreira Kevin Naik Miguel Da Silva Kaveer Harie Ruz Gatawa https://lnkd.in/dAYuQypU

11

Building an agent or RAG system with high accuracy and low cost and latency can be a daunting problem. Between the choice of language models, prompts, depth of reasoning, text splitters, retrievers etc., the options to explore can be innumerous. Hence we build Syftr at DataRobot. Syftr auto-magically searches over the large set of hyperparameters to find the pareto-optimal frontier of configurations given accuracy, cost and latency constraints based on the needs of the agent. Syftr uses Bayesian optimization to efficiently search over the hyperparameter combinations. The set of combinations can be >> 10^23, you might as well give up optimizing this manually. Syftr uses a novel pruning technique that stops early with non-promising search paths. This is how it can efficiently identify the optimal frontier. The best part is syftr is fully open-source. Clone the repo and try it out yourself. It can be configured to support various LLMs, prompt strategies, RAG configs etc. Check it out yourself here: https://lnkd.in/gbPnNW_D DataRobot was a pioneer in AutoML with predictive models, that DNA seeded the idea to build Syftr at DataRobot. Special shoutout Debadeepta Dey for leading Syftr and giving life to this idea. Looking forward to get Syftr natively in DataRobot.

78

1 Comment

Up until recently, most of my research has focused on what AI is doing to software engineering as a discipline: how it changes the process of building software, what it does to our day‑to‑day experience as engineers, and how it shifts where we spend our time when we have an AI assistant sitting beside us in the IDE. But I'm just as intrigued by what it means for the *software itself*. The kinds of systems we can now build. The architectures we reach for. The new constraints we run into and the new classes of problems we have to solve when a non‑deterministic component sits in the middle of everything. All of this has had me musing upon what skills and concepts engineers will actually need to build these systems safely and reliably, so this post is an attempt to lay out a first version of that map: https://lnkd.in/gX2Wr5G5 I'd love to hear what others are seeing out there 👀

44

3 Comments

I tested the new Claude 1 million‑token context window, convinced the larger window would be a game‑changer. The new model changed the game but not in a way I expected. I configured my “software janitor” software to use the new Claude Opus 4.6 model instead of the GPT 5.4.  The software janitor has a context‑gathering agent with read‑only rights that hands off everything it finds to a downstream agent for bug fixes or code review. The flow is bug_fix_context_gatherer → bug_fix_analyzer → bug_fix_developer. Almost immediately the analyzer started acting like an implementer. It tried to write files despite having only read permissions. When it hit the wall it announced,   "It seems write permissions aren't being granted. Let me try a different approach.   Let me try using bash to write the files since that seems to have permissions."   and then attempted to bypass the guardrails. The framework logs captured the rogue actions in real time, forcing me to abort the run. I dug into the prompt and realized the “don’t implement” clause was too weak. I replaced it with a firmer statement: “ANALYSIS ONLY — do NOT modify any files or write any code.” Originally we omitted that line and the system behaved, but after the update the omission allowed the overreach. That stopped some of the strain, but the agent still pushed against its limits. The next fix was to tighten the permission model further, even though we already used granular read/write permissions on the project as a whole, which is why we noticed the behavior in the first place. I gave granular read/write permissions only to the sub directories each agent truly needs and injected the same guardrails into a generic framework that works with multiple models. The result is a zero‑trust system where every capability is explicitly bounded and continuously logged. Now I’m building a dedicated monitoring agent to watch for these anomalies. The non‑deterministic nature of modern agents means we can’t rely on simple statistical checks. Explicit, granular guardrails and continuous monitoring are essential whenever AI agents touch codebases or production systems; trust must be earned, not assumed.

7