We're an AI research lab for applied ML and foundation models. We reverse engineer, interpret and build environments, libraries and frameworks. Our dream is to accelerate ML development with precision and intent for all, enabling safer and better ML across physics, life sciences, industrial, frontier and critical verticals.
Product Now
Trained a talking-avatar LoRA on MiniMax-H3 Ref2VA: one character image plus dialogue becomes a clip with generated speech, lip sync, and light motion. Released with Reactor.
Research Now
Trained world models hub: talking-avatar LoRA on MiniMax-H3, an action-conditioned JEPA for Mario with probe gates, and a 21MB pixel-art world model.
Lab Now
Action-conditioned JEPA for Super Mario Bros with LeJEPA/SIGReg in latent space and probe gates that catch planning failures early.
Lab Now
Geometry beside loss as a controllable training signal — a measurable handle on how representations organize during training.
Partners Now
With the Indian Institute of Technology Patna: fair ride-matching on real NYC taxi data so driver incomes stay even over an evening without losing total earnings.
Research 2026
Open work on mixing models as a practical architecture pattern.
Research Sep 2026
Treat attention as a routing map from one forward pass, keep a small subgraph aimed at the answer, and check it under ablation. On induction and IOI across GPT-2 Small/Medium and Pythia-410M, extracted edges hurt far more than random sets of the same size.
Research Aug 2026
We extract mechanism mounts directly from weight sites by column-tiled SVD, each mount a triple as trigger, write, and strength. Pre-registered A/B/C eval on Gemma-2-2B (WikiText-2): 182/182 GO.
Lab Now
Live, concept-level control of training runs: watch features form mid-run, then steer with data, loss, gradient, and representation interventions. First milestone on fine-tuning with a spurious shortcut.
Lab Now
Cross-study replication of neurodegeneration snRNA-seq findings. One disease, every public dataset, the same pipeline throughout, then a CSF and plasma biomarker shortlist from what holds. Target December 2026.
Research Summer 2026
Activation steering for bias, Chess Agents Simulation, Echoes gossip RPG, and Among Us Agents Simulation.
Partners Now
Interpretable medical ViTs and diffusion models for wound recovery imaging, with major private hospital chains and government hospitals in India.
Lab 2026
Via NVIDIA Inception, Emergent Ventures (GMU), Founders Inc, The Residency, and AWS, Azure & OpenAI Startup Programs. Direction validated with teams at Google DeepMind, Microsoft AI, the Alan Turing Institute, Columbia University, and others.
Partners Now
Building an interpretable VLM that maps text and vision towers on top of OpenCLIP.
Partners Spring 2026
Using interpretability tooling for preference detection analysis in LLMs.
Partners Spring 2026
Training SAEs to sense conflict, ambiguity, and tone and choose the right response. Paper headed to the Interpretability4Discovery workshop.
Product 2025
Open-source CLI and Python SDK for the gap a loss curve does not explain. Inspect, simulate, watch, diff, steer, probe, and eval. Applied interpretability research and tooling.
Research Summer 2026
Agentic ROME with causal tracing and a validation loop that rolls back and retries on failure. Case studies on factuality, bias correction, and censor auditing.
Not sure if Aquin is right for you?
Resources
Policies & licensing
