iamMrDuncan
iamMrDuncan is the public identity and personal project portfolio of Shannon Duncan, who builds open-source, local-first AI tools that are measurable, affordable, and useful on modest hardware. The portfolio focuses on helping people build with AI, improve model and software performance, and learn in public through Hackers in the Loop.
At a Glance
- Local AI enthusiasts
- Solo builders
- Startups
- Enterprises
- +1 more
AI Tools by iamMrDuncan
(1)
Thermite
Open Source AI Autoresearch Agents
Discussions
No discussions yet
Be the first to start a discussion about iamMrDuncan
Latest News
Thermite v0.1.3 released with multi-machine rigs, core/runner architecture, and fixes from campaigns on three DGX Sparks.
Thermite's fourth Qwen3.8-27B-PN3 campaign reported a 16.1% prompt-speed improvement, from 120.5 to 139.2 tokens/second, with decode held near 29.9.
iamMrDuncan highlighted the Qwen 3.8 27B result on one P100, reaching nearly 30 tokens/second.
iamMrDuncan shared that a fully local-model Thermite hill climb produced a 216.3% improvement overnight.
Products & Services
Open-source autoresearch harness for AI agents. It runs hill-climb, goal, and research campaigns on local machines, GPU rigs, Kubernetes, or RunPod while an isolated judge measures changes and keeps only verified wins.
Local AI inference engine and model/weight work aimed at making older GPUs such as the Tesla P100 useful. Its documented Qwen3.8-27B-PN3 result reached 30.03 tokens/second on one P100, with a P100-native PN weight format.
Installable local-first bot app for desktop and phone with persistent chats, file handling, searchable history, local or hosted model connections, skills, and restricted MCP servers; it has no hosted version and requires no account for the machine itself.
Source-backed catalog of free LLM, embedding, speech, image, and research API offers, with dated first-party evidence and separate treatment of ongoing, community, and trial offers.
Market Position
iamMrDuncan positions itself as a local-first, open-source alternative to hosted and opaque AI tooling: projects emphasize running on owned hardware, inexpensive older GPUs, measurable receipts, and inspectable source. Thermite differentiates from ordinary agent coding tools and optimization workflows by keeping measurement and verdict code outside the agents' reach; fxbot differentiates from hosted assistants by running on the user's machine with optional local models and explicitly scoped integrations. The portfolio overlaps conceptually with hosted AI assistants, cloud inference APIs, autonomous coding agents, and autoresearch/benchmarking frameworks, but emphasizes affordability, reproducibility, and user control.
Leadership
Founders
Shannon Duncan
Independent builder and creator of iamMrDuncan's projects; his official biography describes experience starting with mainframes and mainframe databases, then petabyte-scale data engineering including Hadoop clusters with 80,000 cores, product development, and AI systems used by teams and customers. He is also the founder of Hackers in the Loop and works as Principal Architect of Agentic AI at LiveRamp.
Executive Team
Shannon Duncan
Independent creator; founder of Hackers in the Loop
Builder of the iamMrDuncan portfolio, including Thermite, tensorless, fxbot, benchmarking tools, and learning projects. His biography describes a career spanning mainframe databases, large-scale data engineering, product development, and AI systems.
Founding Story
Shannon Duncan says he builds things that solve his own and the community's problems. The current portfolio grew from three goals: build a collaborative place where people can learn and build together, test whether smaller models can outperform larger ones, and make useful AI possible on inexpensive older hardware such as Tesla P100s. Thermite was started to improve inference engines on a P100 and support deeper autoresearch; tensorless was created to make a P100 useful for local AI without spending thousands of dollars on GPUs.
Business Model
Revenue Model
The principal projects are published as open-source or free tools and are designed to run on users' own machines. iamMrDuncan also offers AI integration, agentic-systems, and technical-direction advising; no recurring product subscription or usage-based pricing is stated on the reviewed sources.
Pricing Tiers
The official personal site labels these as current rates and notes that the framing is mostly a joke.
Target Markets
- Local AI enthusiasts
- Solo builders
- Startups
- Enterprises
- Researchers and engineers working with measurable software, model, scientific, or operational problems
- Improving local LLM inference speed and fitting models to available hardware
- Model training efficiency and LLM application quality
- Software performance optimization, migrations, and technology evaluation
- Scientific simulation, mathematical discovery, literature review, and reproduction
- Healthcare, genomics, imaging, clinical text, and FHIR conformance research
- Agriculture, buildings, energy, operations, logistics, robotics, and finance