Hello, I'm

PAWAN.

Data systems & AI governance

Founder of Kapkoti Solution and co-founder of Rigel ERP, building from the hills of Kapkot, Uttarakhand. I make data platforms and the governance that lets AI touch real databases safely: compiled queries, role checks, PII masking and an audit trail.

2,000+sql-sop installs
7Packages shipped
36Merged in drt
Pawan Singh Kapkoti smiling on a train
On-prem first
Now · FounderLive

Kapkoti Solution

Practical AI built around the software a business already uses.

Now · Co-founderProduction

Rigel ERP

GST billing, inventory, manufacturing and a governed AI assistant for Indian businesses.

Always · MaintainerOpen source

sql-steward

An MCP server where an AI agent queries a real database and never writes SQL.

PythonSQLT-SQLTypeScriptdbtPostgreSQLSQL ServerDuckDBBigQuerySnowflakePolarsAirflowPythonSQLT-SQLTypeScriptdbtPostgreSQLSQL ServerDuckDBBigQuerySnowflakePolarsAirflow
FastMCPLangGraphOllamapgvectorChromaDBFastAPIReactNext.jsSupabaseCloudflareDockerGitHub ActionsPower BIFastMCPLangGraphOllamapgvectorChromaDBFastAPIReactNext.jsSupabaseCloudflareDockerGitHub ActionsPower BI
//01 SEC_ABOUT INDEX [01/06]
Who I am

Background & Philosophy

I'm Pawan Singh Kapkoti, a data and AI engineer from Kapkot, Uttarakhand. Contract work on manufacturing data taught me what matters: pipelines that survive Sunday nights, auditors who need answers in plain English, and data that should never leave the building.

That shaped everything I build now, from Rigel ERP to the open-source governance layer around sql-steward. When a tool I rely on has a bug, I try to fix it upstream rather than work around it.

"An AI agent should never be trusted with a database just because it can reach one. Compile the query, check the role, mask the data, keep the receipt."
  1. 01

    Govern before you generate

    Queries are compiled from a semantic layer, not written freely by a model.

  2. 02

    On-prem by default

    Local models and local data. No cloud or API keys needed to get value.

  3. 03

    Every answer leaves a receipt

    Hash-chained audit records show what the agent asked and what it saw.

  4. 04

    Fix it at the root

    Read the source, find the gap, ship the fix upstream with a test.

//02 SEC_JOURNEY INDEX [02/06]
The trail so far

From classroom to company

Every camp on the way up, from a first degree in Noida to building my own company back home in the hills.

2019to 2022 · Noida

BCA, AI and Machine Learning

Amity University. Algorithms, databases and software engineering; a final-year inventory tracker with a JSP front and a MySQL back.

2023to 2024 · Birmingham, UK

MSc Data Analytics

Aston University. Statistical learning, Spark and database systems, with a machine learning dissertation on student outcomes. Plus Microsoft PL-300 (Power BI).

2024to 2025 · UK and India

Data engineer, contracts

Pipelines from source systems into governed warehouses, and BRC/HACCP compliance dashboards with plain-English queries for auditors.

2026March to April

Building in the open

The UK Crime Pipeline, sql-sop on PyPI, and my first merged pull request in drt, where I went on to merge 36.

2026June

The governance layer

agent-blackbox, then sql-steward: tools that let an AI agent use a real database without ever writing SQL.

2026September to now · Kapkot

Kapkoti Solution and Rigel ERP

Founded Kapkoti Solution, an AI engineering practice, and co-founded Rigel ERP. Still learning every day.

//03 SEC_WORK INDEX [03/06]
What I'm building now

Flagship Builds

Two companies and the open-source governance layer that keeps AI honest inside them.

Co-founded product
React · Vite · Supabase · Deno
rigel.kapkotisolution.com →

Rigel ERP: GST-ready business software with a governed AI assistant

End-to-end ERP for Indian businesses: billing, inventory, purchasing, manufacturing, payments and returns in one system. The built-in assistant answers from the company's own books and drafts documents, but a person always presses Save.

522Commits
518DB migrations
28Edge functions
179Test files
7-layer product stack
Click a layer to see what it does.
Founded

Kapkoti Solution: practical AI, built around your business

An independent AI engineering practice in Kapkot. We start with one task that slows a team down, try a small solution, and only build what proves useful. No technical knowledge needed, and you speak directly with me.

Rigel ERPFloorMindElenchusSedno
  • Less paperworkDetails pulled from invoices, forms and documents, ready for your team to check.
  • Answers you can findThe right information from company documents and business records.
  • Tools that work togetherAI features inside the software you already use.
  • A clearer view of the businessBusiness figures turned into useful reports.
AI governance // open source

sql-steward: AI queries a database, never writes SQL

A governed MCP server. Every query is compiled from a semantic layer you control, so there is no run_sql tool to misuse. Listed in the official MCP registry.

  • schema-scout maps 150+ table databases
  • query-warden enforces role access
  • pii-veil masks or blocks PII columns
  • agent-blackbox keeps a hash-chained ledger
Pythonsqlglot
Live demo →
Agents // full stack

Governed Agent Stack

FloorMind turns a plain-English factory question into SQL, then the governance layer lints, gates, masks and audits it. Elenchus runs governed surveys alongside, all behind one Next.js console.

  • One console, 13 pages
  • Every backend proxied server-side
  • SSE streaming answers
Next.jsFastAPI
Code →
Payments // agent safety

The Payments Line

The same idea applied to money: an agent checks policy before it ever reaches a payment rail.

  • steward: an agent spending a sponsor's money
  • pay-warden: budgets, caps and approvals as MCP
  • payoptimize: routes to the rail most likely to accept
MCPPython
Code →
//04 SEC_INDEX INDEX [04/06]
Everything else I've shipped

The Build Index

Pipelines, governance tools and open-source packages. Filter by what interests you.

//05 SEC_UPSTREAM INDEX [05/06]
Giving back

Fixes in tools others rely on

Small, careful pull requests to the open-source projects I use every day, each with the reason it mattered.

~/upstream $ git log --author=pawan --oneline
*mergedsqlglot #7824

Presto and Trino SHA256/512 silently changed hash values in transpilation. Fixed at the root; merged the next morning.

*mergedsqlglot #7832

A follow-up at the maintainer's invitation: BigQuery SHA512 digests with byte-for-byte semantics across six dialects.

*mergedApache Superset #39118

Aligned the dashboard download permission with the explore path, refined after maintainer review.

*mergedSQLFluff #8088

Restored T-SQL single-quote normalization that a lexer patch had dropped for every quoted token.

*mergeddrt #668

sync.mask: hash, redact or truncate PII before load, plus the Mixpanel destination and the connector tutorial.

*mergedOpen Food Facts #13892

Added "groundnut" to the peanut allergen synonyms in a safety-critical dataset used worldwide.

*mergedscanapi #868

Docstrings for spec_evaluator.py, plus a pipx install-path docs PR.

oin reviewSQLMesh #5888

Hex-string surrogate keys for SHA256/512 on Presto and Trino.

oin reviewsops #2262

Stop requiring a matching creation rule when a key is given explicitly.

oin reviewdlt #4026

Resolve a GCP service account against OAuth credentials correctly.

oin reviewdrt #902

The drt serve delivery contract: coalescing, 202 with a run id, pluggable auth.

//06 SEC_CONTACT INDEX [06/06]
Say hello

Let's build something that holds up.

Open to data engineering, on-prem AI and governed-agent work, and always happy to help a business find a practical first step with AI.

pawankapkoti3889@gmail.com

Thank you for scrolling all the way down. Dhanyavaad.