SynapseKit 1.0.0: Python LLM Framework

Asking an LLM to explain your code is easy —if you know exactly what to give it. For a codebase of 50,000+ lines, that's the actual hard problem. Your context window fills up fast, and naive retrieval gives you the function but misses everything it depends on. So I built something that handles this differently. 🔍 Here's what it actually does: You give it a GitHub URL or a local folder path. Then: 1. Parses every Python file using AST to extract functions, classes, parameters, and what each function calls internally 2. Embeds all the code using SentenceTransformers for semantic search 3. Detects your query intent — specific function, full class, or vague question 4. Builds a dependency graph using NetworkX — ask about login() and it automatically pulls in hash_password() and generate_token() because the graph knows what each function calls 5. Sends the structured, relevant context to Gemini and returns a developer-friendly explanation 💡 The key insight that changed everything: The hard part isn't the LLM call. It's building the right context before you make that call. A well-structured 50-line context beats a 5000-line code dump every time. 🛠 Tech stack: Python · AST · SentenceTransformers · NetworkX · Google Gemini API · GitPython 🔗 GitHub: https://lnkd.in/g3J2JCb7 #Python #AI #LLM #MachineLearning #SoftwareEngineering #BuildInPublic

I've been building something for the past few weeks and finally got it to a point I'm happy with. 🚀 A RAG-powered knowledge assistant that lets you ask questions across YouTube videos, PDFs, and websites - and actually tells you where the answer came from. 🔍 Most RAG tutorials stop at "embed ➡️ retrieve ➡️ generate". I wanted to go further and actually think through the tradeoffs: ⚡ MMR over pure similarity search Cosine similarity kept returning near-duplicate chunks from the same paragraph - wasting the entire context window on the same sentence five times. MMR re-ranks candidates to balance relevance and diversity so the model sees actual varied evidence. 🧠 Sliding window memory instead of full history Storing the full chat and sending it on every query means token cost grows linearly with conversation length. Capping at the last 10 turns keeps cost and latency predictable - same performance on query 3 and query 50. 📋 Structured outputs, not raw text Every answer returns a grounded response, key points, a confidence score, follow-up questions, and source citations [S1], [S2] - so nothing is ever a black box. 📊 Measured, not estimated 852ms average MMR retrieval across a live index. Real number, not a guess. 🛠️ Built with Python · LangChain · FAISS · OpenAI · Streamlit ✅ Modular codebase · 13 passing unit tests · dark/light mode UI 🔗 https://lnkd.in/eZyu5TED Would love feedback - especially from anyone who's built something similar and made different tradeoffs. What would you have done differently? 💬 #MachineLearning #LLM #RAG #LangChain #OpenAI #Python #AIEngineering #VectorSearch #BuildInPublic

🚨 Token limits aren’t the real problem — context selection is. While working on LLM pipelines, I kept running into the same trade-off: • Truncate old messages → lose useful context • Send everything → waste tokens and increase cost Neither felt right. So I started experimenting with a different approach: 👉 Treat memory as compression + retrieval What worked surprisingly well: • Older messages → compressed into a short rolling summary (TextRank) • Recent messages → filtered using TF-IDF to keep only what’s relevant • Final prompt → summary + relevant context (not full history) Result: ✔ stays within token limits ✔ preserves important context ✔ reduces unnecessary token usage And the interesting part — this works without heavy infra or embeddings. So instead of asking: “how do I fit everything into the context window?” A better question is: 👉 what actually deserves to be in the context? I packaged this into a small Python library while experimenting. If you're building with LLMs, curious how you're handling memory — truncation, embeddings, or something else? #LLM #AIEngineering #Python #MLOps #RAG #LLMOps

A while ago, I completed an assessment project focused on building and deploying a Flask application with a production mindset. My work comprised of: Flask app architecture Monitoring and observability Security and compliance considerations Infrastructure Provisioning and Deployment best practices It became a classic scenario of; Knowing the skill is important, but speed also matters. This has reinforced my active investment in Agentic AI — and the MCP shift. The Model Context Protocol is looking very much widely adoptable after the LLM + service tools phase. Truth be told, it is tough, but the numerous possibilities help to fuel the drive. Hopefully in the near future peripheral stages of projects are baked, reviewed or assessed with a few models. Feedbacks are welcome: https://lnkd.in/dnh8VdJd #DevOpsProject #Flaskapp #python #AgenticAI #MCP #Bestpractices #Infrasecurity

Last Tuesday, Google Research published TurboQuant — an algorithm that compresses AI memory by 6x with zero accuracy loss. Memory chip stocks dropped within hours. My research engine caught it before most people even heard about it. I’ve been building research_motor — an open-source tool that scans Reddit, GitHub, Twitter/X, and the web, then scores everything it finds across 7 weighted dimensions: • Source trust • Novelty • Actionability • Evidence density • Noise detection • Follow-up potential • Timeliness One command. 140 sources. Ranked digest in seconds. No LLM needed — pure heuristics, zero external dependencies, 88 tests passing. This is part of WinstonRedGuard — a 25-app Python monorepo I built from scratch with V0 discipline: minimal, deterministic, local-first. Now I’m offering these capabilities as services: → AI Research Automation Setup — custom pipelines that scan your industry daily and deliver what matters → Python CLI Tool Development — production-grade command-line tools with strict output contracts → Web Scraping & Data Pipelines — multi-source data collection, scoring, and reporting → GitHub Repo Audit & DevOps — deep technical analysis of any repository with risk detection If you need someone who ships small, tests everything, and trusts data over hype — let’s talk. GitHub: https://lnkd.in/dE_Z89SM #Python #Automation #AI #CLI #WebScraping #DevOps #OpenSource #DataEngineering

I recently had an interview where I was asked how I would build an AI system that can answer questions from 10,000 files. I didn’t have a strong answer. My AI experience was mostly chat history and summarization — not retrieval across a large document set. At the end the interviewer gave me a hint: RAG. So I built it from scratch — a document Q&A API where you upload files and ask questions about them. The workflow: 1. Split documents into chunks 2. Embed each chunk locally using sentence-transformers (free, runs on your machine) 3. Store vectors in PostgreSQL with pgvector 4. Embed the user query 5. Retrieve top 20 candidates via approximate nearest neighbor search 6. Rerank using a cross-encoder model to select the true top 5 7. Generate a grounded answer via Groq API (free tier, Llama 3.1) Built with Python, FastAPI, and containerized with Docker Compose. Used Azure Blob Storage (free tier) for file storage and Groq for inference — the entire stack costs $0 to run. I didn’t get the job. But I turned one weak answer into a project and a much better understanding of retrieval systems. Next time I get that question, I’ll have a real answer. GitHub: https://lnkd.in/e7cDAjdx #RAG #Python #FastAPI #PostgreSQL #LLM #SoftwareEngineering

Day 53 of 100 Days of AI — 🔄 Revision Day Busy day. Didn't build much. But took the time to sit with everything — the newsletter architecture, the Cloudflare setup, the FastAPI backend, the agent pipeline plan. Making sure the full picture is clear before the agent work starts. Sometimes the most productive thing you can do is slow down and make sure you actually understand what you've already built. Agent work resumes tomorrow. Next: Building the synthesis agent — the brain of the newsletter. #100DaysOfAI #BuildInPublic #FastAPI #AIEngineering #Python #Newsletter #SideProject #LangChain

PyMC in the browser? I had to try it myself. But the magic demo from 2022 is broken today. Curious what it takes to revive Bayesian inference in your browser? Read on. I spent my weekend deep-diving into PyMC, PyTensor, and Pyodide, trying to get full Bayesian inference running in the browser again. It was a journey through dependency hell, WASM builds, and open-source realities. I forked PyTensor and made Numba an optional dependency for WASM, enabling PyTensor to build and install in Pyodide. Open-source contributions aren't just about technical fixes—they're about meeting maintainers where they are. My initial Pixi-based dev environment was too disruptive, so I adapted my changes to fit PyTensor's existing mamba workflow. Instead of pushing my preferred tooling, I reworked the PR to respect PyTensor's established environment, focusing on the critical Numba change. The core challenge was making PyTensor (PyMC's computational backend) build for WebAssembly. This meant conditional dependencies, Emscripten setup, and understanding Pyodide's build system. The big win: PyTensor now installs in WASM environments, and PyMC can be imported in the browser. But NUTS—the powerful sampler—doesn't work due to missing JIT support in WASM. Alternative samplers like Metropolis-Hastings might still work for small models, and the roadmap for full browser-based Bayesian inference is clearer now. The infrastructure gap is real, but not insurmountable. If you're curious about the technical details, lessons learned, and the future of Bayesian inference in the browser, check out my full write-up. Would love your thoughts, feedback, or war stories! https://lnkd.in/eKCp7Nru Have you tried running complex Python libraries in the browser? What challenges or surprises did you encounter? #pymc #webassembly #pyodide #opensource #bayesianinference