Open-source AI tools you can try in the browser · ⭐ star what you find useful
| Project | What it does | Try it |
|---|---|---|
| Prompt Lab | Open, model-agnostic prompt optimization: run, score, diagnose, rewrite, repeat. A real Groq run went from 0% to 73%. | Live lab |
| Agora | A public square where AI agents and humans debate in the open; MCP server and A2A listing. | Live debates |
| India AI Index | Open benchmark harness for LLMs on India-specific tasks: Indian languages, Indian documents, cost in INR. | Write-up |
| AgentGuard | Policy engine, scenario suite and simulator for testing AI agents that move money (payments, lending, UPI). | Write-up |
| Ask Your Data | Plain English to SQL and charts on CSV/SQLite. Local-first on DuckDB, works with no API key. | Write-up |
| bounded-autonomy | Policy-as-code governance for AI decisions in regulated finance. | Live console |
Most of what goes wrong with data and AI is not a modelling problem. It is a question of what we choose to measure, what we let machines decide, and who stays accountable.
Head, Business Intelligence Unit · a private-sector bank · Mumbai
15+ years across analytics and AI in banking, financial services, insurance and telecom/media. I write independently about decision intelligence, governance, measurement, data quality, and how intelligent systems earn trust.
I currently head the Business Intelligence Unit at a private-sector bank. Earlier, I headed advanced analytics and digital lending strategy at an NBFC and analytics at a health insurer, with prior roles in media, telecom and technology.
Education: PGDM, XLRI Jamshedpur · PGPMI · Diploma in Advanced Analytics, ISI · FRM Part I (GARP)
Featured · bounded-autonomy · live console
A model that knows when not to act. Policy-as-code governance for AI decisions in banking: stakes tiers, prohibitions, abstention when the data can't be trusted, circuit breakers and a tamper-evident ledger. On 20,000 synthetic bank decisions it automates 54% of them at 18% lower cost than human review of everything, with zero prohibited actions.
Official personal website · ReadSigil · About Shashwat Kumar · Media & interviews
Essays
- Faster cars need better brakes: deciding what a system may do on its own
- When the score becomes the goal: Goodhart's law in the age of optimising agents
- Intelligence without a main character: capability that emerges from orchestration
- Models are only as good as their inputs: make your data boring
- Ideas map →
Code · small, self-contained studies on synthetic data
| churn-propensity-model | Is the clever model worth it? Gradient boosting against a logistic baseline, SHAP drivers, decile lift |
| credit-risk-scorecard | WoE/IV application scorecard with PDO scaling, AUC/KS/Gini and calibration |
english_to_sql.py |
A tiny rule-based English-to-SQL snippet with unit tests |
| bank-analytics-agent | Plain-English question → guarded read-only SQL → answer and chart |
| kpi-dashboard-mis | One set of numbers: KPI dashboard and MIS pack generated end to end from code |
| dpdp-readiness-toolkit | Score a personal-data inventory against India's DPDP Act and Rules |
LinkedIn — Shashwat Kumar · ReadSigil · Media
All demos use synthetic data about fictional companies. Views are my own.