Skip to content
View snehasish01's full-sized avatar

Block or report snehasish01

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
snehasish01/README.md

Snehasish Satpathy. Decision modeling and AI reliability. Drift diffusion sample paths accumulate between two decision boundaries toward a decision point.

Data scientist by profession. Independently, I research decision modeling, evaluate autonomous LLM agents, and build human AI interaction systems.

BS Statistics and Data Science with a Computer Science minor, University of Arizona

Decamint, the hiring and assessment platform I cofounded Snehasish Satpathy on LinkedIn Briefly Aware, my Substack on cognitive science and AI My preprints and data on OSF My ORCID researcher record

Section divider marking a new section

Now

  • Data scientist at John Deere.
  • Cofounder of Decamint, an AI native hiring and assessment platform.
  • Independent Research: cognitive science, human-computer interaction, and human AI symbiosis, focused on how people and AI systems form beliefs, judge their own competence, and decide under uncertainty.

Section divider marking a new section

Research

My work asks one question in two settings: how do reasoners, whether human or artificial, judge their own competence, and how often are those judgments wrong. Two commitments follow from taking that seriously.

Model the failure. Before I trust an LLM agent, a benchmark, or a decision procedure, I want its failure mode taxonomy, its hazard rate across the length of a run, and a measured number in place of an impression. Reliability engineering has spent decades formalizing this for aerospace and manufacturing, and most of the apparatus carries over to autonomous software.

Simulate before you trust. A statistical claim is worth making only after the procedure behind it has been run against a known ground truth often enough to show it meets its guarantees. A confidence interval that misses its stated coverage is wrong, whatever the derivation behind it.

Section divider marking a new section

Selected work

Work What it is Venue and link
Project Bathtub, Phase 0 Failure mode taxonomy for autonomous LLM coding agents. 80 modes across 10 categories, ranked by Risk Priority Number. Zenodo, 2026
sinistra Reproducible high throughput drift diffusion simulation and parameter recovery, with an application to the hand laterality judgement task. Zenodo, 2026
Anticipatory Cognitive Dissonance Preprint. Fires a dissonance inducing prompt while a biased decision is forming rather than after it is committed. OSF
Project ASTRA WebXR study that operationalizes Anticipatory Cognitive Dissonance in a VR decision task. OSF
Decamint AI native hiring and assessment platform. Cofounder. decamint.ai
Briefly Aware Essays on cognitive science, AI, and self knowledge. Substack

Section divider marking a new section

What I work on

AI reliability and evaluation

Reliability engineering applied to autonomous LLM systems: failure mode taxonomies, criticality ranking by Risk Priority Number, and hazard modeling over a long agent run. Alongside it, the statistics to evaluate model behavior without overclaiming, from bootstrap intervals to sequential testing. The connecting idea is calibration. A system that cannot quantify its own uncertainty is asserting, not measuring.

Repositories: bathtub-fmea, eval-power, weibull-bench, claude-hooks-toolkit

Cognitive and decision modeling

Computational accounts of how people decide under uncertainty: drift diffusion models of the speed accuracy tradeoff, behavioral instrumentation of real choices, and interventions that act in real time. The current thread, Anticipatory Cognitive Dissonance, argues that a biased decision is easiest to redirect while it is still forming.

Repositories: sinistra, project-astra

Applied AI and product

Decamint, where I am a cofounder, is an AI native hiring and assessment platform built around structured evaluation of model output. The same interest in making machine behavior observable enough to study runs through the interaction prototypes and the telemetry harnesses.

Link: decamint.ai

Research and writing

I publish preprints and datasets openly, on OSF and Zenodo. Briefly Aware is my Substack on cognitive science, AI, and self knowledge, written for readers outside the field.

Link: Briefly Aware

Section divider marking a new section

Toolkit

Toolkit by area. Languages: Python, Rust, SQL, JavaScript. Statistical computing: NumPy, SciPy, pandas, statsmodels, PyMC. ML and systems: PyTorch, scikit-learn, Pydantic, PyO3, Three.js, WebXR. Research infrastructure: Git, pytest, Parquet, LaTeX, Quarto.

What I reach for first: Python with NumPy and SciPy for analysis, Rust behind PyO3 when a simulation has to be fast, PyMC when the model is Bayesian, pytest and Parquet to keep results reproducible.

Section divider marking a new section

Contact

Pinned Loading

  1. bathtub-fmea bathtub-fmea Public

    FMEA-based failure mode taxonomy (80 items, 10 categories) for autonomous LLM coding agents, ranked by Risk Priority Number

    Python 1

  2. project-astra project-astra Public

    WebXR proof of concept for Anticipatory Cognitive Dissonance — detects forming decision bias from behavioral signals and interrupts with an LLM-generated question before commitment lock-in

    JavaScript

  3. sinistra sinistra Public

    Drift-diffusion model (DDM) simulator with EZ-diffusion and simulation-based parameter recovery — Rust core, Python bindings

    Rust

  4. weibull-bench weibull-bench Public

    Simulation study benchmarking four Weibull parameter estimators (MLE, method-of-moments, probability-plot regression, Bayesian) under censoring — real bias/RMSE and coverage numbers, not a single f…

    Python

  5. eval-power eval-power Public

    Statistical toolkit for comparing LLM outputs — bootstrap CIs, paired McNemar/bootstrap tests, minimum-detectable-effect sizing, and sequential testing with alpha spending. Every method validated b…

    Python

  6. claude-hooks-toolkit claude-hooks-toolkit Public

    Reusable Claude Code hook harness for Project Bathtub — installer CLI, pluggable extractors, JSONL/Parquet sinks, all gated by a 50-check end-to-end test proving real installs and Parquet round-tri…

    Python