Hierarchical & structured reasoning
How can multi-level computation and intermediate representations help models solve problems that are difficult to handle in one pass?
COMPUTER SCIENCE STUDENT · BUILDER
Curious about
how machines
think.
I explore AI reasoning and build software that makes complex systems easier to understand, inspect, and evaluate.
FIELD NOTES / HS—01OPEN SYSTEMS
H
01 / STRUCTURE02 / COMPUTE03 / VERIFYQUESTIONS → EXPERIMENTS → SYSTEMS2026
↳
Long-term focus: building a clearer understanding of reasoning, computation, and intelligent systems.
AI REASONING✳MACHINE LEARNING✳AGENTIC SYSTEMS✳QUANTITATIVE SYSTEMS✳HPC & COMPUTATION✳AI REASONING✳MACHINE LEARNING✳AGENTIC SYSTEMS✳QUANTITATIVE SYSTEMS✳HPC & COMPUTATION
01 ABOUT
I learn by building.
I build to understand.
I'm a computer science student focused on understanding how intelligent systems reason, plan, and solve difficult problems. My primary long-term direction is AI reasoning research — especially hierarchical approaches, deliberate multi-step inference, and the role of computation at inference time.
Alongside that direction, I build software, visual learning environments, and tools for AI-assisted engineering. I treat projects as a way to strengthen fundamentals, test ideas, and make technical concepts easier to inspect. I care about reproducibility, clear scope, and precision about what an experiment actually demonstrates.
[ CURRENT APPROACH ]Understand the problem. Build the smallest useful system. Verify the result. Document the limits.
02 RESEARCH INTERESTS
Questions worth
working on.
These are areas I am learning about and exploring. They are research interests, not claims of completed results.
↗
System 2-inspired inference
When should a model pause, plan, search, or verify — and how should it decide how much additional computation to spend?
Reasoning evaluation
How do we measure correctness, robustness, consistency, and efficiency rather than reward answer length or convincing presentation?
Compute & efficiency
What trade-offs emerge among model size, inference cost, latency, memory, and reasoning quality?
HPC & scientific computing
How can parallel computation and shared compute infrastructure make experimentation more capable, efficient, and reproducible?
FOUNDER & LEAD RESEARCHER · REASONINGAIINDEPENDENT INITIATIVE ↗
FOUNDED AND LED BY HARSH SINGH
ReasoningAI
I founded ReasoningAI as an independent research-engineering initiative exploring inference-time compute, structured reasoning systems, and verifiable agentic reasoning. The work focuses on building practical research infrastructure, evaluation tools, and workflows for investigating how AI systems reason.
Visit reasoningai.tech ↗03 SELECTED SYSTEMS
Research-led work.
Built to be useful.
Flagship systems come first. The additional projects are focused AI tools, agent workflows, and engineering systems. Each link leads to the public repository for implementation details and current scope.
01 / FLAGSHIP SYSTEMSFOUR FEATURED ENTRIES
RESEARCH INFRASTRUCTURE01RAI
ReasoningAI Studio
An evidence-first research toolkit for AI agents and human researchers, combining portable research workflows, reusable prompts, and local Python utilities for transparent, reproducible work.
64 agent skills12 reusable promptsLocal Python utilities
INTERACTIVE MACHINE LEARNING02NN
Neural Networks Playground
A visual-first machine-learning lab for training small neural networks, inspecting activations and gradients, comparing optimisers, and connecting model behaviour to its underlying mathematics.
1–4 hidden layers2–32 units per layerNumPy reference track
QUANTITATIVE & AI RESEARCH03MP
MarketPilot AI
A cross-model trading research toolkit for market discovery, risk checks, scenario replay, walk-forward testing, and portfolio review. Human-led and paper-only, with no live order routing.
17 agent skills11 core workflows6 research modules
AGENTIC ENGINEERING04AG
Agentic Engineering Systems
Two complementary toolkits for reliable agent-assisted engineering: AgentFieldkit supplies evidence-first workflows for software, security, and AI/ML research; LaunchCrew adds spec-first planning, specialist roles, and delivery guardrails for Codex.
25 AgentFieldkit workflows23 LaunchCrew skills10 Codex specialist profiles
02 / ADDITIONAL AI SYSTEMS & TOOLSSPECIALISED PROJECTS
CLAUDE CODECG
Claude Garage
A Claude Code plugin with specialist agents, slash commands, and resumable workflows for planning, implementation, testing, and review.
13 specialist agents · 20 commands
Repository ↗
PORTABLE AGENT SKILLSAW
Agent Workbench
A curated, multi-source skill library and set of portable role profiles for engineering, software delivery, research, and specialist agent workflows.
50 skills · 10 role profiles
Repository ↗
GEMINI CLISG
The Startup — Gemini CLI
A Gemini CLI-native, multi-agent engineering toolkit with spec-driven execution, architecture debate, security review, and FinOps gates.
12 specialist agents · 20 commands
Repository ↗
CREATIVE AGENT WORKFLOWSASG
Agentic Skill Garage
A large portable skill library and local-first Python toolkit for creative AI workflows spanning image, video, audio, 3D, and game-art production.
219 agent skills · Python toolkit
Repository ↗
AI UI QUALITYPX
Pixelcraft
A focused skill library that helps AI agents identify generic, inconsistent, or misleading interface patterns and improve them through evidence-led review.
51 focused skills
Repository ↗
AI & ENGINEERING SIMULATIONKQ
KINETIQ
Portable agent skills for interactive 3D machine explainers, mechanical simulation, physics validation, and offline engineering delivery.
17 simulation skills
Repository ↗
Project counts reflect the linked repositories' published descriptions and may change as the projects evolve. They are not independent performance benchmarks.
REPOSITORY DETAILS & SOURCE CODEBrowse public GitHub repositories ↗
04 TECHNICAL TOOLKIT
Tools are means.
Understanding is the goal.
A working toolkit across software development, machine learning, data, and experimentation. Proficiency varies by language and project; the repositories show the practical context.
Python · Java · C++ · JavaScript · SQL
PyTorch · TensorFlow · NumPy · Pandas · Matplotlib
Git · GitHub · Terminal · Slurm
Reasoning models · Agents · Quant systems · HPC
WORKING PRINCIPLESThink before coding.Keep the system simple.Verify the result.Be honest about evidence.
05 / CONTACT & COLLABORATION
Let's work on
good questions.
I'm interested in thoughtful discussions around AI reasoning, machine learning, model evaluation, agentic engineering, and open-source projects.