# PyTorch Projects at AI Tinkerers

> Canonical HTML: https://aitinkerers.org/technologies/pytorch
> Markdown URL: https://aitinkerers.org/technologies/pytorch.md
> Technology record last updated: 2026-09-18T15:13:53Z
> Generated: 2026-09-21T15:45:29Z

PyTorch is the open-source machine learning framework: it provides a Python-first tensor library with strong GPU acceleration and a dynamic computation graph for building deep neural networks.

PyTorch, developed by Meta AI, is a premier open-source deep learning framework favored in both research and production environments. Its core is a powerful tensor library (like NumPy) optimized for GPU acceleration, delivering 50x or greater speedups for complex computations. The key differentiator is its 'Pythonic' design and dynamic computation graph (eager execution), which allows for rapid prototyping and simplified debugging compared to static-graph frameworks. Leveraging its Autograd system for automatic differentiation, practitioners build and train models for computer vision and NLP; major companies like Tesla (Autopilot) and Microsoft utilize PyTorch for critical AI applications.

- Official technology site: https://pytorch.org
- Public AI Tinkerers demos and talks: 273
- Result page: 1 of 12

## Recent Public Talks and Demos

### [The Goldilocks zone between autopilot and a blank file](https://minneapolis-saint-paul.aitinkerers.org/talks/rsvp_T02lAwAfSBo)

Most ML tooling makes you leave the tool to get to the next layer. Templates are on Hugging Face, helpers are scattered across GitHub, compute is somewhere else again, and the first week goes to stitching instead of modeling. The usual fix is to point a code model at the gaps, which works badly here for a specific reason: bad ML code does not crash, it runs. You watch the loss curve descend for four hours, pay for the GPUs, and find out the helper was quietly wrong the whole time. So ACI sits between full autonomy and a rigid template. Sub-autonomous, in the sense that the assistance is retrieval and ranking rather than generation. Two ranking models I trained order verified project templates and PyTorch helper functions out of a library curated by hand, currently 178 core blocks and 7,719 across 16 extension packs. A language model does write the one line plain English summary on each block and powers the help chat. It never writes code and never touches the ranking. The converter is what I actually want to show. Block graph to real PyTorch and back, fully deterministic, surviving nested blocks across two serialization formats. Hand edit the emitted script and it parses back into blocks. It is a compiler, not a template you cannot touch. Deployed at aimlse.org, so I will just run it. Real login, open a project, go template to blocks to script, then edit the script and watch the blocks change to match. Then throw a cell at a sandboxed PyTorch kernel container through a Redis queue and watch the output stream back with the colors intact. Three processes on one EC2 box under Docker Compose. run_api on 8000 serving the frontends, auth on 8001, job processor eating the queue. App code baked into the image, static assets bind mounted, which I did for reasons I am still not sure were correct. The editor is open on the site with no account needed, so anyone can pull it up during the demo and try to break the compiler while I am talking.

- Event context: AI Tinkerers Minneapolis-Saint Paul — July Meetup — 2026-07-29 — Minneapolis Saint Paul
- Public talk page: https://minneapolis-saint-paul.aitinkerers.org/talks/rsvp_T02lAwAfSBo

### [Robot Picks Grapes Based on Spatial World Model](https://seattle.aitinkerers.org/talks/rsvp_hwjgpFnwm2M)

We built a new model that is based on the human brain. Right now it can navigate and reason to play Pac Man but we are working on connecting it to a robot and having it pick grapes.

- Event context: AI Tinkerers Seattle Summer Bash — 2026-07-29 — Seattle
- Public talk page: https://seattle.aitinkerers.org/talks/rsvp_hwjgpFnwm2M

### [The loss curve lied: catching hidden safety drift inside the fine-tuning, automated with an agent!](https://zurich.aitinkerers.org/talks/rsvp_Vnt244xXFE0)

We built a fine-tuning where a Claude-Code based agent that helps founders and researchers fine-tune models on their datasets, and get more insights than just the training and eval losses. In particular, we also show the model’s regress on safety and general capability, metrics that otherwise would be overlooked. For example, we give it a medical instruction dataset; the agent asks a few multiple-choice questions and proposes concepts like "recommends unsafe or contraindicated actions," "downplays red-flag symptoms that need urgent care," and "implies clinical authority it doesn't have," then writes the run config. Then, for each concept, it fits a single direction in the model's activations (difference-of-means, following Persona Vectors, Chen et al. 2025, arXiv:2507.21509) that “encodes” that concept, and reads that direction off every checkpoint during LoRA training. Live we can show the agent conversation proposing the concepts, as how they drift alongside per-checkpoint evals. This is part of an ongoing research agenda with a collaboration of researchers from ETH Zurich, ETH AI Center and Agentic Systems Lab. We formulated our research into an agentic harness and live demo in an anthropic hackathon last week, and want to share more about it! Also, we imagine this as an open-source effort.

- Event context: AI Tinkerers Zürich: July Build Night with Swisscom Ventures — 2026-07-01 — Zürich
- Public talk page: https://zurich.aitinkerers.org/talks/rsvp_Vnt244xXFE0

### [stress-testing autonomous research agents](https://zurich.aitinkerers.org/talks/rsvp_N7RI28XXj20)

I built a minimal autoresearch loop for coding agents, and I test the agent-environment behaviour during standard and broken evaluations. I will show how the system is structured as a research sandbox. I will walk through the live workflow, and I will show when and how the different evals I tested changed the agent-environment behaviour.

- Event context: AI Tinkerers Zürich: July Build Night with Swisscom Ventures — 2026-07-01 — Zürich
- Public talk page: https://zurich.aitinkerers.org/talks/rsvp_N7RI28XXj20

### [Training a small world model into a superhuman player at competitive pokemon](https://toronto.aitinkerers.org/talks/rsvp_UxzGmf_2Ipo)

World models are the hot new "it" in AI, and I will demo a simple world model capable of superman performance in 1v1 competitive pokemon play in pokemon showdown. I will walk over the following points 1) what is competitive pokemon and why is it harder than Go. I will explain why pokemon team selection is the largest factor in success, followed by strategic pokemon switches. 2) the world model encoder, predictor and decoder architecture and training pipeline. A deeper analysis will go over some of the training logs in weights and biases, the tokenizer and embedding interpretability 3) The world model bot will be available for players to challenge on pokemonshowdown.com

- Event context: AI Tinkerers Toronto - June 2026 - with Relay Financial x PostHog — 2026-06-30 — Toronto
- Public talk page: https://toronto.aitinkerers.org/talks/rsvp_UxzGmf_2Ipo

### [Your Webcam Knows the Geometry - Real-Time Relighting & Single-Shot Novel Views](https://lausanne.aitinkerers.org/talks/rsvp_cMMuinG2WV4)

Two single-view inverse-rendering systems that reconstruct scene geometry from one camera and re-render it under conditions never captured — one changes the light, the other changes the viewpoint, both fast enough to ship. The Relighting part decomposes a live webcam frame into geometry + material (normals/albedo/shading) and re-renders it under any lighting in real time (~24 FPS), showcasing a variety of lighting conditions (multiple colored lights, orbiting and rainbow-rotating lights) The Novel View Synthesis builds a 3D representation of the scene using Gaussian Splatting and renders it from a different angle. It currently runs at 5 FPS but we are hoping to get closer to real time soon.

- Event context: AI Tinkerers Lausanne June 2026 Meetup — 2026-06-25 — Lausanne
- Public talk page: https://lausanne.aitinkerers.org/talks/rsvp_cMMuinG2WV4

### [A Guardrail for the Hardest Conversations: (Bilingual) Youth Crisis Detection](https://montreal.aitinkerers.org/talks/rsvp_noPwfE2FLcI)

A stateful, multi-turn input guardrail that screens an entire sensitive youth mental health conversation arc (not just the latest input) to catch crises that build up gradually ("slow drift"), the failure mode where each turn looks benign but the cumulative trajectory is high-risk. Live, I'll walk through the actual system: the two-stage stack (fine-tuned mmBERT classifier → Cohere c4ai chain-of-thought judge), the 5-question reasoning prompt that made the difference, and the evaluation harness output on a hidden validation set (F1 0.899, recall 0.954 at ~1.16s/sample). I'll show the architecture diagram, the prompt engineering and the red-team CSVs that trained the classifier.

- Event context: AI Tinkerers Montreal - June Demo Meetup — 2026-06-17 — Montreal
- Public talk page: https://montreal.aitinkerers.org/talks/rsvp_noPwfE2FLcI

### [Real-time segmentation of teeth on dental X-rays using a mobile phone.](https://columbus.aitinkerers.org/talks/rsvp_NnfSI8ddrD0)

The project involves the creation of a custom trained AI model that detects and segments teeth on dental X-rays using the real-time video feed of an Android phone. I will discuss the tools used and provide a working demo of the app or a video of it working.

- Event context: AI Tinkerers - Columbus June Meetup — 2026-06-01 — Columbus
- Public talk page: https://columbus.aitinkerers.org/talks/rsvp_NnfSI8ddrD0

### [Transformer-Diffusion model for molecular battery material generation](https://dubai.aitinkerers.org/talks/rsvp_SK0KgeEQ2Rc)

I built Simer Energy, an end-to-end generative AI pipeline that uses a hybrid Transformer-Diffusion architecture to autonomously design, physically relax and thermodynamically validate novel battery materials from scratch. For the demo, I will execute a terminal-based run of the by inputting strict material constraints (e.g., elemental bounds for a cobalt-free transition metal oxide, target stoichiometry, and symmetry parameters) into a fine-tuned model. I’ll show how the Transformer maps these constraints into a discretespace groups and passes it as a conditioning vector to a diffusion model. You will see the model denoise the 3D spatial atomic coordinates, followed by the critical "zero-strain" and then passing the raw generated structure directly into a Universal Machine Learning Force Field (CHGNet) to instantly relax the atomic coordinates, bypassing days of expensive DFT calculations and then run the relaxed structure through ALIGNN to validate its Energy Above Hull, resulting in a mathematically viable file generated in under three minutes.

- Event context: AI Tinkerers Dubai - May Demo Day — 2026-05-23 — Dubai
- Public talk page: https://dubai.aitinkerers.org/talks/rsvp_SK0KgeEQ2Rc

### [Speak mk1: A multimodal mamba-attention hybrid model for speech therapy](https://dubai.aitinkerers.org/talks/rsvp_f9s0FLsNTT0)

I am currently building a mamba-attention hybrid encoder and decoder from scratch for speech therapy, not fine-tuning, not wrapping an API, but training every component end to end on my RTX 4060 laptop. The system has three components I am building in parallel: The first is a custom Mamba SSM-based audio encoder trained on LibriSpeech with multi-task phonological heads for voicing, manner, place of articulation, and correctness detection. The encoder uses a BLIP-2-style Q-Former to bridge audio representations into the language model. The second is a video pipeline using MediaPipe FaceLandmarker to isolate and analyze oral region motion frame by frame, extracting articulatory features like tongue tip position, mouth opening geometry, and lip protrusion in real time from a standard webcam. The third component is SpeakMK1LLM, which serves as the project’s core reasoning engine. To balance cutting-edge architectural research with deployment stability for the current phase, we have developed two iterations of this model: The Hybrid Prototype: A custom 70M parameter Mamba-attention hybrid model designed for high-efficiency sequence modeling. It was trained on a rigorous four-stage curriculum: general pretraining on Tinystories, domain adaptation on CHILDES, clinical knowledge injection from pubmed central, and final instruction tuning. The Deployment Model: A fine-tuned Gemma 3n E4B, which leverages the same four-stage curriculum to adapt its large-scale reasoning capabilities to the specific nuances of Speech-Language Pathology (SLP). The entire stack requires only a webcam and a microphone, which is the point — making clinical-quality articulation feedback accessible without specialized hardware or a therapist present in the room. For the demo I would show the live pipeline taking a child's speech, flagging a specific phoneme error, and generating a graded corrective prompt in real time, alongside the clinician dashboard that logs session data for remote review.

- Event context: AI Tinkerers Dubai - May Demo Day — 2026-05-23 — Dubai
- Public talk page: https://dubai.aitinkerers.org/talks/rsvp_f9s0FLsNTT0

### [Privacy vs. Accuracy in Federated Learning](https://doha.aitinkerers.org/talks/rsvp_C8C7h2ua7_s)

A benchmarking system that tests federated learning algorithms FedNova, FedAvg, and FedProx under differential privacy constraints to find the optimal balance between model accuracy and data privacy.

- Event context: AI Tinkerers – Doha Inaugural Meetup — 2026-05-18 — Doha
- Public talk page: https://doha.aitinkerers.org/talks/rsvp_C8C7h2ua7_s

### [Porting Segment Anything: SAM2 in ONNX/C++, SAM3 in thin scripts](https://lausanne.aitinkerers.org/talks/rsvp_HLLk1mziQvk)

I built a portable interactive segmentation stack around Segment Anything: a C++/ONNX Runtime wrapper for SAM2 plus thin SAM3 image/video demo scripts that run on Windows CUDA or macOS Apple Silicon. In the demo I’ll show prompt-driven segmentation and video propagation, then pop the hood on how I split the model/runtime pieces so the same workflow runs outside research notebooks.

- Event context: AI Tinkerers Lausanne April 2026 Meetup — 2026-04-30 — Lausanne
- Public talk page: https://lausanne.aitinkerers.org/talks/rsvp_HLLk1mziQvk

### [MakoraGenerate - AI Agent for Optimizing GPU Code](https://poland.aitinkerers.org/talks/rsvp_bSNP53ASxRc)

MakoraGenerate writes highly performant GPU kernels, low-level code that is the backbone of modern AI compute infrastructure.

- Event context: AI Tinkerers Poland - Meetup in Gdańsk #1 — 2026-04-23 — Poland
- Public talk page: https://poland.aitinkerers.org/talks/rsvp_bSNP53ASxRc

### [eomer: An Agentic Operating System for Data Science](https://singapore.aitinkerers.org/talks/rsvp_2G9vgEuwDJc)

eomer is an agentic operating system for data science that fine-tunes tabular foundation models on proprietary data to deliver production-ready forecasting, regression, and classification pipelines. In this demo, I’ll show how we go from raw enterprise data (CSV / DB) → automated feature handling → model selection & fine-tuning → calibrated forecasts with uncertainty → deployable outputs (API/UI) in a single workflow, without bespoke ML engineering.

- Event context: AI Tinkerers Singapore: The Agentic Future & Dev/Eng Workflows — 2026-04-21 — Singapore
- Public talk page: https://singapore.aitinkerers.org/talks/rsvp_2G9vgEuwDJc

### [MoE parameters, MoE problems: visualizing Mixture of Experts Routing Layers](https://manchester-nh.aitinkerers.org/talks/rsvp_vIT84Fso8c4)

I built a real-time neural telemetry engine designed to intercept and visualize the internal gating decisions of a Mixture of Experts (MoE) model at a deeper level. The project utilizes a PyTorch forward hook to capture raw 40-dimensional routing weights from Layer 20 of an IBM Granite 3.0 model. By running this entirely on a laptop, the demo provides a live "neural heartbeat" that proves how sparse activation can achieve high-performance reasoning without the latency or computational waste of a cloud-scale cluster.

- Event context: AI Tinkerers Manchester (Bedford), NH - April 2026 Meetup — 2026-04-15 — Manchester NH
- Public talk page: https://manchester-nh.aitinkerers.org/talks/rsvp_vIT84Fso8c4

### [Shop Talk](https://st-louis.aitinkerers.org/talks/rsvp_HCVQpW0TgEw)

I built ShopTalk, a multimodal shopping assistant that lets users search for products using text, images, or a combination of both, delivering more human-like and relevant results. In the demo, users can upload an image, type a query, or do both, and the system uses BLIP to generate contextual captions and CLIP to embed both text and images into a shared vector space, retrieving and ranking the most relevant products through a Redis-backed similarity search with late fusion scoring.

- Event context: AI Tinkerers St. Louis: April 2026 Meetup — 2026-04-14 — St. Louis
- Public talk page: https://st-louis.aitinkerers.org/talks/rsvp_HCVQpW0TgEw

### [Replacing Analysts in commodity trading](https://zurich.aitinkerers.org/talks/rsvp_i8MtV8raWAE)

Informatiom system built to deliver what traders want in LNG / Gas speculative decisions. - Visualize Fundamentals - Machine Learning models - Automate processes (anomalies alerts, analyze competitors, check auctions) - Data architecture automation - News embeddings and semantic analyzes (check what moves the market) - Monitoring users for automated feedback

- Event context: AI Tinkerers Zurich April 9th — 2026-04-09 — Zürich
- Public talk page: https://zurich.aitinkerers.org/talks/rsvp_i8MtV8raWAE

### [From scratch, single file implementation of a scaled-down version of MuZero for CartPole/Atari(WIP)](https://zurich.aitinkerers.org/talks/rsvp_hNEGEsBZehc)

I'm reimplementing reinforcement learning algorithms from scratch, most recently MuZero. I want to understand what comes after LLMs. Models learning purely from experience of interacting with the environment seem interesting, see for example David Silver's recent start-up announcement. I want to show that it is possible to reimplement core parts of these algorithms in a single file.

- Event context: AI Tinkerers Zurich April 9th — 2026-04-09 — Zürich
- Public talk page: https://zurich.aitinkerers.org/talks/rsvp_hNEGEsBZehc

### [Let's meet in latent biology space client side for a beverage!](https://miami.aitinkerers.org/talks/rsvp_b4DYTOqH6fM)

Demo and discuss CytoVerse (https://www.biorxiv.org/content/10.64898/2026.01.29.702554v1) - a client side web app that runs a biology foundation model (single cell rna-seq embedding) client side so scientists don't have to upload data, can get immediate feedback on an experiment and share their results in latent space preserving their data for publication.

- Event context: AI Tinkerers Miami: Spring Demos at The Lab — 2026-03-25 — Miami
- Public talk page: https://miami.aitinkerers.org/talks/rsvp_b4DYTOqH6fM

### [Your Brand Translator](https://paris.aitinkerers.org/talks/rsvp__5FOk0bFKFc)

A personal branch is a must for today's professionals and entrepreneurs. But these same people don't have time to do it themselves and services cost thousands. So we build an agent, - based on OpenClaw for easy interaction via WhatsApp, Telegram and Signal - with a ton of image and video editing tool: IG filters, automatic clipping, deep image generation, caption generation, video from image generation, video modifications, music generation, application of any HF-based model - and the ability to schedule the posts via buffer Share your expertise and experience effortlessly.

- Event context: High-Performance Local AI Development: Kick-off ThinkStation PGX — 2026-03-17 — Paris
- Public talk page: https://paris.aitinkerers.org/talks/rsvp__5FOk0bFKFc

### [Turning Claude Code into a Self-Improving Agentic Development Platform](https://orange-county.aitinkerers.org/talks/rsvp_iO2rjPLhMOE)

Every AI coding session is a chain-of-thought reasoning trace — step-by-step problem solving with verifiable outcomes. BashGym captures these traces and uses them to train a reasoning language model with the same techniques behind frontier RLMs: GRPO for reinforcement learning, RLVR for verifiable reward signals from test results, and distillation to transfer reasoning from a large teacher into a small local model. The result is a personal RLM trained on how you actually think through code — your conventions, your repos, your patterns.

- Event context: AI Tinkerers OC - March 11, 2026 Meetup — 2026-03-12 — Orange County
- Public talk page: https://orange-county.aitinkerers.org/talks/rsvp_iO2rjPLhMOE

### [Words to World, AI learning strategies for building world models.](https://san-diego.aitinkerers.org/talks/rsvp_mnJaK0Bd4l8)

Can a text prompt (or reference art) build a physically real world, no splats, no diffusion, but full physics? Once you have this, can a robot or biocomputer navigate it? Full demo of the system running live (local / remote hybrid) How it is all plumbed and built. The system uses a variety of learning strategies including self-modifying code and rules to both iteratively improve a world model and, iteratively improve the generation of that model AND, iteratively improve the improvement of that model (3 level hierarchy). It uses a large range of optimising strategies to avoid uncanny valley effects and attempt to produce cinematically realistic images.

- Event context: AI Tinkerers San Diego: February Meetup at Google — 2026-02-27 — San Diego
- Public talk page: https://san-diego.aitinkerers.org/talks/rsvp_mnJaK0Bd4l8

### [Meta Modeling for drug discovery.](https://dc.aitinkerers.org/talks/rsvp_SBMb6DF9kMI)

I am training a meta model for ml based chemical binding prediction on open data. I already have about 1500 hundred fine tunes of binding predictions and the meta model will be used to predict the viability of future fine tunes. I am also investigating applying symmetry breaking to align binding symmetries with chemical point groups. This could tune the meta model to specific chemistries. If done this would be a new architecture. I currently have a simple web interface to show the results of my training runs and will be adapting it to output the results of the meta model. It is running locally but I may be able to have it open by the meeting. A lot of my work has been created with assistance from chatgpt and codex.

- Event context: AI Tinkerers - DC Metro Meetup - February 24th, 2026 — 2026-02-24 — DC
- Public talk page: https://dc.aitinkerers.org/talks/rsvp_SBMb6DF9kMI

### [Exercise Posture Tracking Using Computer Vision](https://raleigh.aitinkerers.org/talks/rsvp_QACg5zj28MM)

I've created an unsupervised computer vision algorithm to detect back rounding during deadlift. In short, I've demonstrated that I can measure the curvature of the back in an unsupervised way, and that a rounded back produces a different signal from a flat back. I've used off-the-shelf pose tracking and segmentation models to measure the back curvature. This could help guide learning of proper form during deadlift and other exercises.

- Event context: AI Tinkerers Raleigh Meetup — February 11, 2026 — 2026-02-11 — Raleigh
- Public talk page: https://raleigh.aitinkerers.org/talks/rsvp_QACg5zj28MM

## Related Technologies

- [GPT-4](https://aitinkerers.org/technologies/gpt-4) ([Markdown](https://aitinkerers.org/technologies/gpt-4.md)) — 529 public demos
- [LangChain](https://aitinkerers.org/technologies/langchain) ([Markdown](https://aitinkerers.org/technologies/langchain.md)) — 445 public demos
- [OpenAI API](https://aitinkerers.org/technologies/openai-api) ([Markdown](https://aitinkerers.org/technologies/openai-api.md)) — 520 public demos
- [ONNX](https://aitinkerers.org/technologies/onnx) ([Markdown](https://aitinkerers.org/technologies/onnx.md)) — 83 public demos
- [TensorFlow](https://aitinkerers.org/technologies/tensorflow) ([Markdown](https://aitinkerers.org/technologies/tensorflow.md)) — 90 public demos
- [scikit-learn](https://aitinkerers.org/technologies/scikit-learn) ([Markdown](https://aitinkerers.org/technologies/scikit-learn.md)) — 84 public demos
- [BERT](https://aitinkerers.org/technologies/bert) ([Markdown](https://aitinkerers.org/technologies/bert.md)) — 179 public demos
- [Keras](https://aitinkerers.org/technologies/keras) ([Markdown](https://aitinkerers.org/technologies/keras.md)) — 74 public demos
- [GPT-3](https://aitinkerers.org/technologies/gpt-3) ([Markdown](https://aitinkerers.org/technologies/gpt-3.md)) — 191 public demos
- [Python](https://aitinkerers.org/technologies/python) ([Markdown](https://aitinkerers.org/technologies/python.md)) — 662 public demos
- [Transformers](https://aitinkerers.org/technologies/transformers) ([Markdown](https://aitinkerers.org/technologies/transformers.md)) — 148 public demos
- [Llama-2](https://aitinkerers.org/technologies/llama-2) ([Markdown](https://aitinkerers.org/technologies/llama-2.md)) — 227 public demos
- [Claude-3](https://aitinkerers.org/technologies/claude-3) ([Markdown](https://aitinkerers.org/technologies/claude-3.md)) — 110 public demos
- [BLOOM](https://aitinkerers.org/technologies/bloom) ([Markdown](https://aitinkerers.org/technologies/bloom.md)) — 115 public demos
- [PaLM 2](https://aitinkerers.org/technologies/palm-2) ([Markdown](https://aitinkerers.org/technologies/palm-2.md)) — 116 public demos
- [RoBERTa](https://aitinkerers.org/technologies/roberta) ([Markdown](https://aitinkerers.org/technologies/roberta.md)) — 118 public demos
- [FastAPI](https://aitinkerers.org/technologies/fastapi) ([Markdown](https://aitinkerers.org/technologies/fastapi.md)) — 181 public demos
- [Docker](https://aitinkerers.org/technologies/docker) ([Markdown](https://aitinkerers.org/technologies/docker.md)) — 147 public demos

## More Results

- Next: https://aitinkerers.org/technologies/pytorch.md?page=2