I build the layer between a raw LLM and something a business can actually rely on — retrieval that cites its sources instead of hallucinating them, and agents that fail loudly instead of silently.
🌐 Portfolio: sid-surange.github.io · 🔭 Currently building: Briefcast — an AI research briefing agent running in production for ~$8/month
- 🧠 What I do — 10 years in software, last 2 focused on AI Engineering — shipping LLM-powered products to production, not just notebooks
- 🤖 Right now — deep in agentic dev tooling (Claude Code, Cursor), RAG that's actually evaluated (dedup, ranking, citations, before/after evals), and local quantized LLM setups
- 🏠 How I build — locally first (LM Studio, quantized Llama/Granite), cloud APIs only once it's proven
- ✍️ Outside the day job — writing on Medium about the parts of GenAI that are underdocumented, and shipping experiments to Hugging Face Spaces
| Category | Tools |
|---|---|
| Languages & AI |      |
| Agentic Tooling |    |
| Backend & Infra |       |
Automated AI research briefing agent — monitors Google AI, DeepMind, OpenAI, Anthropic, arXiv and more, then delivers a curated daily digest to Telegram with follow-up Q&A over a 14-day rolling knowledge base.
Stack: FastAPI · PostgreSQL + pgvector · LangChain LCEL · OpenRouter (Gemini, Claude) · APScheduler · Railway Highlights: dual-layer deduplication (SHA-256 + cosine similarity), tiered source ranking, RAG answers with citations, ~$8/month to run in production
Scans official and top community Claude Code repos for agents, skills, commands, and hooks, then lets you interactively install what you want straight into ~/.claude/ — with license tracking and attribution built in.
Stack: Python · Claude Code ecosystem tooling Highlights: 8 curated sources (official + community) with per-repo license classification, unified scanning across heterogeneous repo layouts, interactive installer
A team-ops layer for Cursor — versioned rules, agent skills, and git guardrails that roll out across an entire engineering team in minutes, not sprints.
Stack: Shell · Cursor rules/skills · CI hooks Highlights: 16 versioned skills (core + community tiers), 7 enforcement rule packs, 3 automated guardrail hooks (git-guard, migration-guard, license-gatekeeper), CODEOWNERS + CONTRIBUTING + SECURITY policy baked in
HR-focused resume analysis tool that runs entirely on local LLMs — no API keys required. Extracts structured data from PDFs, flags missing sections, generates tailored interview questions, and visualizes resume content.
Stack: FastAPI · Gradio · PyTorch · IBM Docling · LM Studio (quantized Llama 3.1/3.2, IBM Granite) Highlights: fully offline, 8-bit quantized model support, spell-check analysis, word cloud generation
Monorepo of production-grade experiments. PageSense: Chrome extension + FastAPI/Qdrant backend for semantic search over your browsing history. AgentForge: deployed agentic app with web search and image generation.
Stack: FastAPI · Qdrant · Gradio · smolagents · LlamaIndex · JavaScript (extension)
- Briefcast: How I Built a Personal AI Intelligence Agent That Reads the Entire AI Ecosystem — For ~$10/Month
- What's new with OpenAI's gpt-4o-mini
- Deciphering the power of Vision Language Models
- AgentForge: A simple AI Agent with Web Search and Image Generation
- ProGAN, StyleGAN, StyleGAN2: Exploring NVIDIA's breakthroughs
Oracle Certified Generative AI Professional · Google Cloud Professional Data Engineer
