https://space.gimhara.com An engineering publication by Kushan Manahara exploring AI systems, machine learning, agents, and software architecture. en hi@gimhara.com (Kushan Manahara) https://space.gimhara.com/articles/anthrpic-project-glasswing https://space.gimhara.com/articles/anthrpic-project-glasswing Tue, 26 May 2026 00:00:00 GMT In one month a restricted Claude model scanned over 1,000 open-source projects and isolated more than 10,000 high or critical severity bugs. Research https://space.gimhara.com/articles/engineering-council-sri-lanka https://space.gimhara.com/articles/engineering-council-sri-lanka Tue, 01 Apr 2025 00:00:00 GMT Receiving official registration under the ECSL Act No. 4 of 2017. What statutory engineering recognition means for software and systems practitioners. Findings https://space.gimhara.com/articles/mcp-vs-acp https://space.gimhara.com/articles/mcp-vs-acp Tue, 25 Feb 2025 00:00:00 GMT The consolidation under open standards settled the protocol question. MCP solves tool access; A2A solves agent collaboration. Here is how they stack in production. Systems https://space.gimhara.com/articles/docker-mcp https://space.gimhara.com/articles/docker-mcp Mon, 24 Feb 2025 00:00:00 GMT Fifty tool definitions crammed into the context window costs tokens, degrades reasoning, and risks hallucinations. Docker's containerized MCP Gateway and dynamic discovery flip the model. Engineering https://space.gimhara.com/articles/my-mcp https://space.gimhara.com/articles/my-mcp Sat, 22 Feb 2025 00:00:00 GMT A standard way to hand a model your files, your tools, and your prompts, without custom glue code holding it all together. Systems https://space.gimhara.com/articles/xcodeai https://space.gimhara.com/articles/xcodeai Tue, 18 Feb 2025 00:00:00 GMT How Apple built predictive code completion directly into Xcode 16 using the Apple Neural Engine. Zero cloud roundtrips, offline privacy, and Swift specialization. Engineering https://space.gimhara.com/articles/cag-vs-rag https://space.gimhara.com/articles/cag-vs-rag Wed, 12 Feb 2025 00:00:00 GMT RAG searches for information on every turn. CAG pre-computes the KV cache. Where Time-to-First-Token and conversational latency matter, caching flips the architecture. Inference https://space.gimhara.com/articles/llm-think https://space.gimhara.com/articles/llm-think Wed, 05 Feb 2025 00:00:00 GMT Tokens, vectors, attention, and next-token prediction. What is actually happening underneath something that feels like a conversation. Research https://space.gimhara.com/articles/lpu-vs-gpu https://space.gimhara.com/articles/lpu-vs-gpu Thu, 30 Jan 2025 00:00:00 GMT Language Processing Units are purpose-built for language models rather than general compute. The specialization shows up in where the weights live. Inference https://space.gimhara.com/articles/la-fires-vs-chatpgt https://space.gimhara.com/articles/la-fires-vs-chatpgt Sun, 12 Jan 2025 00:00:00 GMT ChatGPT did not cause the wildfires. The conversation that formed around the claim is still worth having. Evaluation https://space.gimhara.com/articles/openai-own-chip https://space.gimhara.com/articles/openai-own-chip Sun, 05 Jan 2025 00:00:00 GMT There are not enough GPUs to go around, so the company behind ChatGPT is looking at making the hardware itself. Systems https://space.gimhara.com/articles/llmhg https://space.gimhara.com/articles/llmhg Wed, 25 Dec 2024 00:00:00 GMT Gemma is capable and open source. Hugging Face is what makes getting to it straightforward. Engineering https://space.gimhara.com/articles/llm-finetuning https://space.gimhara.com/articles/llm-finetuning Wed, 18 Dec 2024 00:00:00 GMT Full parameter training requires clusters of H100s. Parameter-Efficient Fine-Tuning (PEFT) with low-rank matrix decomposition brings specialized adaptation to a single GPU. Experiments https://space.gimhara.com/articles/nvidia-nim https://space.gimhara.com/articles/nvidia-nim Tue, 10 Dec 2024 00:00:00 GMT Running self-hosted LLMs used to mean stitching vLLM, Triton, and CUDA drivers by hand. NVIDIA NIM packages TensorRT-LLM in standardized, production-ready OCI containers. Inference https://space.gimhara.com/articles/aws-bedrock https://space.gimhara.com/articles/aws-bedrock Mon, 02 Dec 2024 00:00:00 GMT A web-based prototyping environment with foundation models, knowledge bases, agents, and guardrails. Public preview, two regions. Engineering https://space.gimhara.com/articles/gemini-vector-database https://space.gimhara.com/articles/gemini-vector-database Mon, 25 Nov 2024 00:00:00 GMT Pinecone Serverless for sub-50ms vector retrieval, Google's Gemini for grounded synthesis. A complete end-to-end Python implementation with metadata filtering. Engineering https://space.gimhara.com/articles/gamini-api https://space.gimhara.com/articles/gamini-api Mon, 18 Nov 2024 00:00:00 GMT Massive context windows, native multimodality, and structured schema outputs: how developers can build production systems on Google AI Studio. Engineering https://space.gimhara.com/articles/npm-vs-yarn-vs-pnpm https://space.gimhara.com/articles/npm-vs-yarn-vs-pnpm Sun, 10 Nov 2024 00:00:00 GMT They all install packages. Where they diverge is the shape of node_modules, and that difference has consequences for disk space and for correctness. Engineering https://space.gimhara.com/articles/crewai-agents https://space.gimhara.com/articles/crewai-agents Sun, 03 Nov 2024 00:00:00 GMT Moving beyond single-agent prompt loops: orchestrating multi-agent collaboration with specialized roles, tools, and sequential or hierarchical processes. Engineering https://space.gimhara.com/articles/how-rag-works https://space.gimhara.com/articles/how-rag-works Mon, 28 Oct 2024 00:00:00 GMT A vector database, cosine similarity, and an LLM context injection. The mathematical and code mechanics of grounding language models. Evaluation https://space.gimhara.com/articles/explaiable-ai https://space.gimhara.com/articles/explaiable-ai Tue, 22 Oct 2024 00:00:00 GMT High test accuracy is meaningless if a model exploits spurious correlations. How cooperative game theory and local linear surrogates bring mathematical interpretability to deep learning. Research https://space.gimhara.com/articles/spacex-reverse-rocket https://space.gimhara.com/articles/spacex-reverse-rocket Mon, 14 Oct 2024 00:00:00 GMT On 13 October 2024, the Super Heavy booster came back to the pad and was caught mid-air by a pair of mechanical arms, on the first attempt. Experiments https://space.gimhara.com/articles/logistic-regression https://space.gimhara.com/articles/logistic-regression Sat, 05 Oct 2024 00:00:00 GMT Despite the name, it is a classification algorithm. It models the probability of a binary outcome from one or more predictor variables. Research https://space.gimhara.com/articles/linear-regression https://space.gimhara.com/articles/linear-regression Sat, 28 Sep 2024 00:00:00 GMT A supervised algorithm that models the relationship between your features and a target, on the assumption that the relationship is a straight line. Research https://space.gimhara.com/articles/windows-vs-linux https://space.gimhara.com/articles/windows-vs-linux Fri, 20 Sep 2024 00:00:00 GMT I started on Windows 10 and 11 and thought I had everything I needed. Then the projects got complex enough to show me otherwise. Systems https://space.gimhara.com/articles/go https://space.gimhara.com/articles/go Thu, 12 Sep 2024 00:00:00 GMT How Rob Pike and Ken Thompson's deliberate refusal of language complexity turned Go into the undisputed operating system of the modern cloud. Engineering https://space.gimhara.com/articles/c https://space.gimhara.com/articles/c Thu, 05 Sep 2024 00:00:00 GMT From llama.cpp and TensorRT to low-latency game engines, modern C++ remains the undisputed champion of bare-metal control and zero-cost abstractions. Systems https://space.gimhara.com/articles/sir-don-bradman https://space.gimhara.com/articles/sir-don-bradman Wed, 28 Aug 2024 00:00:00 GMT Don Bradman hit only six sixes in his entire Test career and finished with an average of 99.94. The mathematics of eliminating failure modes in sport and software. Findings https://space.gimhara.com/articles/industry-visit https://space.gimhara.com/articles/industry-visit Thu, 15 Aug 2024 00:00:00 GMT Inside Creative Software with the Peradeniya Computer Engineering team. Observing how enterprise architecture, CI/CD, and agile engineering operate at global scale. Findings https://space.gimhara.com/articles/my-graduate https://space.gimhara.com/articles/my-graduate Thu, 01 Aug 2024 00:00:00 GMT Reflections on completing my B.Sc. (Hons) in Engineering at Peradeniya: four years of systems programming, hardware-software co-design, and lifelong engineering friendships. Findings https://space.gimhara.com/articles/apple-gen-ai-sub-domain https://space.gimhara.com/articles/apple-gen-ai-sub-domain Sat, 25 May 2024 00:00:00 GMT What the registration of genai.apple.com signaled: on-device 3B foundation models, the Apple Neural Engine, and the verifiable security of Private Cloud Compute. Research https://space.gimhara.com/articles/fyp https://space.gimhara.com/articles/fyp Mon, 20 May 2024 00:00:00 GMT Developing an end-to-end long-read transcriptomics pipeline using Oxford Nanopore sequencing: basecalling, splice-aware alignment, isoform discovery, and differential expression. Findings https://space.gimhara.com/articles/my-journey https://space.gimhara.com/articles/my-journey Fri, 15 Mar 2024 00:00:00 GMT My uncle gave me a PC when I was eleven. Almost every part of it eventually broke, and repairing it became the way I learned. Findings https://space.gimhara.com/articles/django-and-python-web-development https://space.gimhara.com/articles/django-and-python-web-development Sun, 12 Mar 2023 00:00:00 GMT From an empty Linux machine to a working Django app: environment, project, models, views, templates, and URLs. Engineering https://space.gimhara.com/articles/stop-conda-from-automatically-activating-the-base-environment https://space.gimhara.com/articles/stop-conda-from-automatically-activating-the-base-environment Sun, 12 Mar 2023 00:00:00 GMT Conda was slowing my machine down, so I stopped it from activating the base environment in every new shell. Engineering https://space.gimhara.com/articles/pip-install-pipenv-issue-solved https://space.gimhara.com/articles/pip-install-pipenv-issue-solved Mon, 06 Mar 2023 00:00:00 GMT Python 3.11 refused the install because the environment is externally managed. That is a deliberate protection, and there are three ways past it. Engineering https://space.gimhara.com/articles/install-my-sql-in-kali-linux https://space.gimhara.com/articles/install-my-sql-in-kali-linux Sat, 04 Mar 2023 00:00:00 GMT Kali ships with MariaDB, and it conflicts with MySQL. Here is the sequence that finally worked for me. Systems https://space.gimhara.com/articles/kali-linux-update-warninig https://space.gimhara.com/articles/kali-linux-update-warninig Sat, 04 Mar 2023 00:00:00 GMT Copying the keyring into trusted.gpg.d silences the warning. Scoping the key to the one repository that needs it is the fix apt actually wants. Systems https://space.gimhara.com/articles/programing-environment-os https://space.gimhara.com/articles/programing-environment-os Fri, 17 Feb 2023 00:00:00 GMT There is no single best answer. It depends on your preferences, your languages, and the tools you work with. Systems https://space.gimhara.com/articles/intro-to-linux https://space.gimhara.com/articles/intro-to-linux Fri, 17 Feb 2023 00:00:00 GMT Open source, stable, and built around the command line. What makes Linux a good place to develop. Systems