I build and evaluate experimental systems for structured reasoning, knowledge-graph discovery, verification, and AI agents. My portfolio focuses on turning architectural ideas into inspectable implementations and testable research questions.
Research question: Can explicit structure—claims, relationships, assumptions, contradictions, and evidence—help AI systems generate better hypotheses and verify them more reliably?
| Project | Focus | Current evidence / status |
|---|---|---|
| Geometric Engine Intelligence | Knowledge-graph representation, unexplored-edge ranking, contradiction-driven discovery | Research prototype with a documented data model, scoring concepts, and evaluation-harness foundation. Claims of knowledge or scientific novelty require independent evaluation. |
| HONet | Continual-learning architecture and an executable observation-to-hypothesis discovery loop | Repository documents demos and unit-test commands. Performance claims should be interpreted in light of independently reproduced benchmark results. |
| Aurora / VINCI AI | Surfacing assumptions, missing context, contradictions, and links across conversations | README describes a deterministic Python prototype, regression tests, CI, and a runnable demo. |
| Meta-Attention Reasoning Toolkit | Experimental meta-attention, confidence estimation, and self-critique | Developer-preview proposal; its README still describes planned SDK components. |
| SUPHAI Model | Multi-agent coordination and adaptive AI system architecture | Architecture-oriented project; benchmark against single-agent and simpler orchestration baselines is a key next step. |
| Synthetic Reason — Collective Intelligence | Symbolic memory, agent collaboration, and consensus | Repository documents a FastAPI-based prototype and optional external service integrations; deployment and reliability need separate verification. |
Observe → Represent claims and evidence → Map relationships
→ Detect assumptions and contradictions
→ Generate hypotheses → Design discriminating tests
→ Evaluate against baselines → Replicate or revise
- Python and applied AI/ML engineering
- Knowledge graphs and structured reasoning
- AI agents and orchestration architectures
- Evaluation harnesses, reproducibility, and falsification
- Geometry, symbolic representation, and cross-domain discovery
These are areas of active project work, not a claim of expertise in every framework or method.
- Runnable examples over unsupported claims.
- Strong baselines and ablations over impressive-sounding demos.
- Reproducible negative results as well as positive findings.
- Clear separation between algorithmic novelty, knowledge novelty, and validated scientific discovery.
- Documentation that makes it easy for another engineer to reproduce a result.
I welcome technically rigorous collaboration on evaluation, graph reasoning, agent systems, reproducibility, and independent replication. Start with a project's README and open an issue with a concrete question, counterexample, or proposed experiment.
Profile: github.com/Anurag1 · Repositories: Browse projects
Project descriptions summarize repository documentation; they are not independent verification of all advertised capabilities. Benchmark superiority and scientific discovery should only be claimed after controlled, reproducible evaluation.