Nripanka Das — AI product leader working across agent systems, energy, climate, and product outcomes.
I turn questions about safety, reliability, and user value into working tools with decisions that can be replayed, inspected, and challenged.
These independent projects show the problem, the design choices, and the proof. Their demos use synthetic inputs; the results below demonstrate software behavior, not customer adoption or production impact.
Help shape the next release: try a short task and give feedback. Critical reviews and blocked attempts are welcome.
Problem: a successful agent demo does not explain which tool calls should be denied, require review, or leave an audit record.
Built: a policy simulator with deny-overrides rules, explicit approval states, secret redaction, and a verifiable audit chain.
Demonstrated: an 11-entry attack fixture produces four allowed calls, five denials, one approval requirement, and one protocol error; its audit verifies. Experimental policy lab with local MCP transports; it does not execute upstream tools or authenticate approvals.
Read the case study · Open the demo · Explore the code
Problem: a dispatch policy can look efficient while leaving demand unmet or requesting actions beyond equipment limits.
Built: a replayable microgrid simulator that records requested and feasible actions, then reports reliability, cost, emissions, and safety together.
Demonstrated: the 48-hour synthetic scenario serves 98.72% of demand while exposing 34.49 kWh of unserved energy and two constraint violations. The trace makes those shortcomings visible. This is an educational simulator.
Read the case study · Open the demo · Explore the code
Problem: time spent in a product can rise because a workflow is confusing.
Built: an executable reference for my Value Density framework, combining outcome quality, time, friction, regret, and harm in an auditable report.
Demonstrated: 16 synthetic invoice sessions compare two flows while keeping completion rates, scoring components, uncertainty, and guardrails visible. The comparison illustrates the measurement method; it establishes no causal effect on real users.
Read the case study · Open the demo · Explore the code
PatchGym turns Git history into coding-agent tasks; RunMirror records and replays agent behavior; Climate Evidence Bench checks quantitative climate answers.
Browse the project index · Current health review · Release and validation history
I work across AI product leadership, energy, and supply chain. This is an independent lab with no employer branding or implied endorsement. Development uses AI assistance, disclosed in each featured repository.
My working principles: preserve human takeover, leave inspectable evidence, expose the components of a score, and state the limits of a benchmark.
Professional context and writing on LinkedIn. For a reproducible bug or a useful integration, open an issue in the relevant repository.