Sign in

Thorin

@tmtabor.io
2.4K followers 737 following 206 posts

Staff Software Engineer specializing in agents, RAG and MCP applications. Nineteen years of full-stack engineering. Open source developer. Building and writing at tmtabor.io

PostsRepliesMedia
Thorin @tmtabor.io · 28/09/2026
I've updated my portfolio with my latest agentic AI, including a Python agent template, a local multi-agent content drafter via Ollama, and an LLM issue triager for GitHub. Check out what I'm building: tmtabor.io/projects #AgenticAI #Python
tmtabor.io
Projects — Thorin Tabor
Projects by Thorin Tabor — agentic AI systems, research software and open source developer tooling.
141
Thorin @tmtabor.io · 25/09/2026
Jev is a new model gaining traction. It's fast and cheap, designed to classify unstructured data instead of generating text. This shift back to non-generative methods is important for certain NLP tasks. #LLMs #AgenticAI #NLP
340
Thorin @tmtabor.io · 24/09/2026
Our agent pipeline kept producing malformed JSON. Fine-tuning fixed it. Here's what that process actually looked like. #AIEngineering
230
Thorin @tmtabor.io · 22/09/2026
The OSS Notifier Agent takes a list of projects you care about. On schedule, it scans for new issues and filters for tickets that are small, clearly described, unblocked, and ideally labeled as good first issues for contributors. #AgenticAI #MultiAgentSystems #AIAgents
github.com
GitHub - tmtabor/oss-notifier-agent: LLM-triaged good-first-issue digest for GitHub repos, delivered by email. Runs on GitHub Actions, no server required.
LLM-triaged good-first-issue digest for GitHub repos, delivered by email. Runs on GitHub Actions, no server required. - tmtabor/oss-notifier-agent
230
Thorin @tmtabor.io · 17/09/2026
Claude Code is a better agent harness than GitHub Copilot and offers better hand-holding than Cursor. But for the senior dev who knows exactly what they are doing? I am not sure it takes the crown. #AgenticWorkflows
560
Thorin @tmtabor.io · 15/09/2026
With GenePattern Copilot we found that a non-augmented model initially beat our RAG assistant. The failure was naive chunking splitting Q&A pairs. The fix was preprocessing with LLM fact extraction, ensuring every vector DB row was an atomic statement. Read the blog post below. #RAG #AgenticAI
tmtabor.io
A hallucinated module, a backfiring RAG pipeline and the MCP server that fixed it — Thorin Tabor
How an eval suite exposed a RAG pipeline that hurt GenePattern Copilot more than it helped, and what finally fixed it.
370
Thorin @tmtabor.io · 14/09/2026
Model Context Protocol (MCP) has no settled answer to a foundational question: Who are you, and what are you allowed to do? #MCP
430
Thorin @tmtabor.io · 12/09/2026
Nemotron 3.5 Lightning is a vivid local model. It consistently invents specific, compelling detail. In controlled throughput tests, it out-generated Gemma 4 on raw tokens/sec, and delivered twice the distinct ideas on an open-ended brainstorm. #AgenticAI #LLMs #AIAgents
Nemotron ProseMuse ProseGemma Prose
120
Thorin @tmtabor.io · 11/09/2026
My job-hunting agent sent me a daily email listing new job postings, noting where I'd be a fit and where I'd be a stretch. It saved hours otherwise spent crawling job boards, but it doesn't write applications or resumes. That must be done deliberately by you. #AgenticAI #AIAgents #LLMs​
github.com
GitHub - tmtabor/job-agent: Daily job-scanning agent: fetches postings from multiple sources, filters/scores against a candidate profile with an LLM, emails a ranked digest. Runs on GitHub Actions.
Daily job-scanning agent: fetches postings from multiple sources, filters/scores against a candidate profile with an LLM, emails a ranked digest. Runs on GitHub Actions. - tmtabor/job-agent
330
Thorin @tmtabor.io · 09/09/2026
You thought debugging an agentic workflow was hard? Try running a "human computer program" with twenty 4-year-olds. It turns out kids are a great test case for clear instruction tuning. #SoftwareEngineering
230
Thorin @tmtabor.io · 08/09/2026
Two years ago I built a #RAG pipeline for GenePattern Copilot, an agentic assistant for genomics analysis. It used #LangChain and had retrieval over documentation and a decade's worth of help forum data. At the time, that was unusual.​ #AgenticAI #BuildInPublic
460
Thorin @tmtabor.io · 05/09/2026
Nemotron 3.5 Lightning passed tool-use grounding tests cleanly. But in open-ended writing, it hallucinated resource ownership details outside the established context. Passing a tool-use test doesn't mean the model is grounded everywhere. The risk moves. #AgenticAI #LLMs #RAG
250
Thorin @tmtabor.io · 03/09/2026
Building a production-grade AI chatbot requires moving past the standard HTTP request-response cycle. While REST APIs work for static data, managing multi-turn agentic updates -- tool calls, hidden reasoning, token streaming -- demands an event-driven architecture. #SystemDesign
830
Thorin @tmtabor.io · 01/09/2026
Open source is fundamentally valuable. The response to careless automation isn't eliminating automation; it's building systems that respect people's time and attention. Responsible scaling requires smart tooling. #AgenticAI #MultiAgentSystems #LLMs
github.com
GitHub - tmtabor/oss-notifier-agent: LLM-triaged good-first-issue digest for GitHub repos, delivered by email. Runs on GitHub Actions, no server required.
LLM-triaged good-first-issue digest for GitHub repos, delivered by email. Runs on GitHub Actions, no server required. - tmtabor/oss-notifier-agent
450
Thorin @tmtabor.io · 31/08/2026
Most MCP servers are just thin wrappers. Here's what it actually takes to make one an agent can use reliably: #MCP
410
Thorin @tmtabor.io · 29/08/2026
Muse Glimmer is not a replacement for latency-sensitive tasks. But for output quality and reasoning depth over raw speed (e.g., planning, content generation, multi-round review), it's the most interesting open-weight model I've tested this year. #AgenticAI #LLMs #AIAgents
Comparison of Muse Glimmer versus Gemma 4 and Nemotron Lightning
320
Thorin @tmtabor.io · 28/08/2026
Are any of you working on vertical-specific agent harnesses, or are you sticking to general-purpose tools for now? I’m curious to see if other fields are seeing this level of specialized integration.​ #AgenticWorkflows #AIAgents​
350
Thorin @tmtabor.io · 27/08/2026
It is a humbling experience to spend months building an agent harness only to have a startup come along and blow your work out of the water.​ #BuildInPublic #AIAgents
350
Thorin @tmtabor.io · 26/08/2026
Embeddings capture meaning, but they do not guarantee precision. When your application logic demands perfect accuracy, you must pair semantic search with structural and explicit ID signals to build a robust system. #AgenticAI #RAG #RAGTips #LLMs
350
Thorin @tmtabor.io · 25/08/2026
I spent more time fighting LangChain than building the actual AI layer for GenePattern Copilot. So I switched to Pydantic AI. #PydanticAI​
pydantic.dev
Pydantic AI
451
Thorin @tmtabor.io · 24/08/2026
Frontier models still lead on novel problems and long horizon planning. But for much of agent work, the gap is smaller than expected. Rebuilding the scaffolding closes more of the distance than upgrading the model does. #AgenticAI #AIAgents #LLMs
570
Thorin @tmtabor.io · 22/08/2026
I put Meta's Muse Glimmer through the same agentic tasks as my daily-driver model. It lost on speed, but it won on almost everything else. Testing new open-weight models is a good exercise in understanding the trade-offs between raw throughput and output fidelity. #AgenticAI #AIAgents #LLMs
360
Thorin @tmtabor.io · 21/08/2026
Frontier models often ignore scaffolding edges. Weaker models find every soft spot in the pipeline. The fix: Validate every tool call against a schema, check result shapes, and ensure shorter steps fail independently. This significantly improved system robustness. #AgenticAI #MultiAgentSystems #LLMs
450
Thorin @tmtabor.io · 20/08/2026
DiffusionGemma generates 256 tokens per step and can rewrite any of them mid-generation. #GenerativeAI
DiffusionGemma
240
Thorin @tmtabor.io · 19/08/2026
Chunking must respect boundaries. Data prep needs to be structure-aware; IDs cannot separate from their blocks. Implement re-ranking that boosts precision using exact ID matches. Test with adversarial near-duplicates—this validation is critical for robust systems. #RAG #RAGTips #AgenticAI
230
Thorin @tmtabor.io · 18/08/2026
I already knew much of what's in this book and I still learned something from every chapter. That's AI Engineering by Chip Huyen. #BookReview
AI Engineering
230
Thorin @tmtabor.io · 17/08/2026
Human interviewers detect subtle signals, an awkward pause or body language, and follow up on them in real time. A scripted model cannot adjust to what is missing; it simply advances through the predetermined script regardless of conversational lossiness. #AgenticAI #LLMs #MultiAgentSystems
130
Thorin @tmtabor.io · 13/08/2026
Every AI agent needs config, logging, evals, and guardrails. I built a template so you don't have to. #AIAgents
agent-template repo
330
Thorin @tmtabor.io · 12/08/2026
Pair vector search with metadata layers for ID retrieval. This explicit matching handles identifiers alongside similarity scores. Hybrid retrieval combines dense vectors with sparse search, letting an exact match like 16-B override a near-duplicate embedding score. #RAG #RAGTips #AgenticAI #LLMs
130
Thorin @tmtabor.io · 11/08/2026
Scaling an MCP server is mostly a solved problem. Scaling what's behind it is where things get interesting. #MCP
MCP
240
Thorin @tmtabor.io · 10/08/2026
I asked an AI interviewer about engineering culture. The response was basically a read-aloud HR brochure. This highlights how easily scripted models fail when faced with nuanced questions that require pattern recognition beyond boilerplate corporate language. #LLMs #AgenticAI #MultiAgentSystems
1320
Thorin @tmtabor.io · 09/08/2026
Sound familiar? #AI #Meme
GenAI
271
Thorin @tmtabor.io · 06/08/2026
Pydantic AI's new capabilities feature quietly solves the biggest problem in agent code: reuse. Here's what stood out to me in v2: #AIAgents
Pydantic AI
330
Thorin @tmtabor.io · 05/08/2026
Embeddings capture meaning but fail on precision. Do not rely on vector scores for exact IDs. Semantically similar identifiers (like 16-A/16-B) confuse retrieval as near-duplicates. Solving this requires augmenting semantic search with explicit structural signals. #RAG #RAGTips #AgenticAI #LLMs
120
Thorin @tmtabor.io · 04/08/2026
AI agents don't need heavier data pipelines. They need smarter, structured context. The newly proposed Open Knowledge Format (OKF) uses progressive disclosure to give LLMs what they need, when they need it, without blowing up token costs. I'm optimistic about this spec. #AIAgents
330
Thorin @tmtabor.io · 03/08/2026
This is a truly useful tool for ensuring that your website is agent-ready, accessible, secure and optimized for SEO. 👉 specification.website/ #AgenticAI #WebDev​
specification.website
The Website Specification
A platform-agnostic, full specification of the technical features a good website should have. Built in the open under an MIT licence.
150
Thorin @tmtabor.io · 02/08/2026
Interesting thoughts on cognitive surrender and implementing intentional friction in AI workflows. Talk by Kathy Baxter at the #AgenticAISummit. #AgenticAI
130
Thorin @tmtabor.io · 02/08/2026
Some great tools coming out of Google for identifying deepfakes, tracing the provenance of images and ensuring that AI bridges communities. Talk by Chris Bregler at the #AgenticAISummit. #AgenticAI
110
Thorin @tmtabor.io · 02/08/2026
Ring true? #AI #AgenticAI #Meme
AI in demos
241
Thorin @tmtabor.io · 01/08/2026
Great talk on agentic tooling and self-direction by Alex Graveley at the #AgenticAISummit.
110
Thorin @tmtabor.io · 01/08/2026
"Science is not getting faster with AI, it's getting slower." --Mengdi Wang at the #AgenticAISummit #AgenticAI
231
Thorin @tmtabor.io · 01/08/2026
Understanding model cognition with Eric Ho at the #AgenticAISummit. Important for reverse-engineering what's going on in a model's weights during inference. #AgenticAI
110
Thorin @tmtabor.io · 01/08/2026
James Zou using #AIAgents to model therapeutics at the #AgenticAISummit.
020
Thorin @tmtabor.io · 01/08/2026
Markus J. Buehler talks AI agents in scientific discovery. #AgenticAISummit #AgenticAI
120
Thorin @tmtabor.io · 01/08/2026
AI-accelerated chip tradeoffs being discussed by Peter DeSantis at the #AgenticAISummit. It's cool. Chip design is something I don't know much about in depth. #AgenticAI
110
Thorin @tmtabor.io · 01/08/2026
"We're at a critical point in AI safety." --Dawn Song #AgenticAISummit #AgenticAI
020
Thorin @tmtabor.io · 01/08/2026
Agentic AI Summit starting at Berkeley. Live posting it today! (skeeting...? Yeah, no.) #AgenticAI
110
Thorin @tmtabor.io · 31/07/2026
The most common architectural failure in modern AI development is the urge to replace logic with prompts. While LLMs are highly capable, shoehorning them into tasks better served by deterministic methods creates a stack that is slow, expensive and unpredictable. #AIArchitecture
220
Thorin @tmtabor.io · 30/07/2026
The secret to a high-performing agent isn't more instructions, it's more signal. If you want to escape context rot; it pays to stop using generic skills and start forging your own library. #AgentSkills​
110
Thorin @tmtabor.io · 29/07/2026
A good agent skill isn't just a list of instructions; it is the distillation of a senior engineer's methodology. If you are building on the agentskills.io spec, curation is the only way to save your context window. #AgenticWorkflows​ #AgentSkills
agentskills.io
Agent Skills Overview - Agent Skills
A standardized way to give AI agents new capabilities and expertise.
120