Aquin Labs
Tldr: We're an AI lab for applied ML and data science. We reverse engineer, interpret, and build universal predictors that research groups and ML teams can tune to their data. Our dream is applied prediction that compounds over time for all, enabling labs to understand what models learn and teams to run classification, annotation, and decision work across tables and media, growing one base with every new workload towards better and safer intelligence across industries.
Aleph is a universal predictor for applied ML and data science. Built for every vertical that needs a right answer: tables, media, annotation, and decisions, and for models that keep learning as the work arrives. One base. Any modality. Any task. Fine-tune into anything and keep getting better. Not a chat assistant. Schema-bound outputs for programs: probs, labels, boxes, scores, with confidence so you escalate the unsure. Built from scratch for precision and intent, deliberately, not by guessing.
- Our work:
- Aleph: universal predictor for applied ML and data science; one base, any modality, continual fit / fine-tune, schema-bound outputs
- axi: coding and research agent with an auto RL (arl) harness and self-learning mission loop
- aq: framework for foundation models and applied ML; YAML map + SDK/CLI, jobs/fleet
- Applied interpretability research and tooling
- Tile-SVD: finding usable weight mechanisms. arXiv:2608.06969
- Attention as a Routing Graph: live circuit extraction. arXiv:2609.25285
- Trained world models, LLM behaviour sims, weight editor for LLMs and MoM: Mixture of Models
- Collaborations:
- AEIA Lab: training SAEs to sense conflict, ambiguity, and tone and choose the right response; paper headed to the Interpretability4Discovery workshop
- KAIST: using interpretability tooling for preference detection analysis in LLMs
- LAION: building an interpretable VLM that maps text and vision towers on top of OpenCLIP
- Zyogen: interpretable medical ViTs and diffusion models for wound recovery imaging
- Indian Institute of Technology Madras: transformers whose position encodings are made of learnable, input-responsive Hopf oscillators
- Indian Institute of Technology Madras: algorithm to plan delivery routes for millions of stops that finishes much faster, while keeping the same quality of routes
- Indian Institute of Technology Patna: fair ride-matching algorithm on taxi data so driver incomes stay even without losing total earnings
- Frontier labs: SAE training and applied interpretability
- $100K in compute and funds via NVIDIA Inception, Emergent Ventures (GMU), Founders Inc, The Residency, and AWS, Azure & OpenAI Startup Programs
- Current focus:
- Cross-study replication of single-nucleus transcriptomic findings in neurodegeneration
- Steering what models learn: live, concept-level control of training runs
- Superposition geometry signal: geometry beside loss as a controllable training signal
- Team: Ash Manvi, Neha Sarda, Paul Cannon, Sambhav Choudhary, Samreena Tajreen, Sachin Sharma
- Socials: aquin.app, aquinlabs.com, aquin@aquin.app, X, YouTube, LinkedIn, Hugging Face, GitHub
