Building systems for reliable
machine intelligence.
Blackdrome AI Labs is an AI research and engineering lab working across agent infrastructure, reinforcement-learning environments, benchmarking, model behaviour, reasoning, and simulation.
We build experiments, tools, and research systems around problems that emerge as AI moves from generating answers to taking actions.
Research that shapes the lab.
Noma: an open, calibrated decision model
An open-weight decision model for agents and pipelines. It reads a state and typed questions and returns calibrated probabilities, an abstain signal, and an uncertainty estimate in one forward pass, in about 16 ms, with no generated text.
Causal Information Flow in Transformer Residual Streams
Causal experiments using activation patching, intervention, probing, and ablation to study how specialist attention heads interact through GPT-2's shared residual stream.
LEMMA
A neuro-symbolic mathematical reasoning system combining transformer-guided proposal generation, Monte Carlo Tree Search, symbolic rules, and verifiable state transitions.
What we investigate.
Agentic Systems
Long-horizon reasoning, tool use, stateful interaction, and autonomous behaviour.
Model Intelligence
Internal representations, mechanistic interpretability, model intervention, and behavioural analysis.
Reasoning & Verification
Search, symbolic constraints, structured reasoning, and verifiable computation.
Simulation & Learning
Reinforcement-learning environments, synthetic experience, robotics, and adaptive learning systems.
Benchmarking & Red Teaming
Agent benchmarks, reward-hacking analysis, adversarial evaluations, and repeatable tests for unsafe optimisation and deceptive behaviour.
Built, measured, and documented.
Noma: An Open System One Decision Model from a Sliced Decoder and Trained Heads
A preprint that documents how Noma is built and tests which parts of its design matter, with open weights and code. Atul Saxena and Divyanshi Sharma, October 2026.
Activation-space interventions for frozen language models
A controlled SAE-guided intervention framework for testing whether edits inside a frozen language model produce reliable behavioural change, including evidence when they do not.
Photorealistic Simulation and Synthetic VLA Infrastructure
A photorealistic aerial-agent environment that joins drone dynamics, reconstructed scenes, waypoint supervision, and synthetic trajectories for reproducible embodied-AI research.
We study what happens after models become capable.
As AI systems become more autonomous, intelligence becomes a systems problem. Models must reason across time, interact with tools, respond to changing state, recover from errors, and learn from feedback.
Notes, experiments, and technical writing.
Research notes and technical writing coming soon.
The people behind Blackdrome.
Interested in the same problems?
We are open to conversations with researchers, engineers, AI teams, and collaborators working on difficult problems in machine intelligence.