Research toward systems that can discover.
Kernellum studies systems that can reason under uncertainty, investigate open questions, test hypotheses, and contribute to reliable new knowledge.
Research Mission
Our work spans machine reasoning, scientific discovery, AI systems & architecture, and evaluation & reliability.
Discovery requires more than generation.
We study systems that can formulate questions, run structured investigations, use evidence to revise beliefs, and expose their own failure modes.
The goal is not to produce more answers. It is to build systems capable of producing reliable new knowledge.
Research Programs
Four programs, one objective: move AI from response generation toward reliable discovery.
- 01
Machine Reasoning
Methods for structured inference, abstraction, search, and reasoning under uncertainty.
- 02
Scientific Discovery
Systems that generate hypotheses, design tests, update from evidence, and pursue open-ended investigations.
- 03
AI Systems & Architecture
AI-native systems and hardware research, including architecture search, physical-design feedback, and efficient computing.
- 04
Evaluation & Reliability
Evaluation methods for reasoning, calibration, robustness, reproducibility, and failure under distribution shift.
Research publications
Technical reports
Methods, measured results, and the artifacts behind them. Read the full reports or follow the evidence back to the repository.
TR-002
21 September 2026 / Technical report
Project SIMILARITY: Learning Routed Timing Laws for Architecture Search
A causal topology intervention isolates operand-distribution topology as an architecture variable, freezes a routed timing law, and tests a bounded two-candidate selection policy on unseen deployment routes.
TR-003
20 September 2026 / Technical report
Closed-Loop Physical-Design Search
Final-route observations are folded back into architecture selection under a fixed implementation budget, comparing active acquisition against an equal-budget random control.
TR-001
17 September 2026 / Technical report
Kernellum Compiler and Physical-Evidence Pipeline
A system-level record connecting quantized model-to-RTL generation, analytical search, synthesis calibration, final-route measurement, closed-loop physical feedback, and the K2 board-validation handoff.
Evidence & reproducibility
21 September 2026 / Evidence dossier
Kernellum Evidence Dossier
A claim-to-artifact record connecting the compiler, Machine Reasoning, Project SIMILARITY, machine-readable results, and the K2 validation handoff into one diligence path.
AI Systems & Architecture
Current program: from routed evidence to physical measurement
K2 is board-ready, with generated bitstreams, UART transport, cycle counters, and frozen board variants. No physical-board latency, power, or energy result is claimed yet.
A separate 72-route B-local prospective study has also been frozen before routing, with preregistered electrical, area, throughput, and evidence gates.