Sanitation workflows
Delivered FSM inbox and workflow improvements, module unbundling, TQM rollout work, localisation, responsive UI fixes, and configuration across SANITATION, digit-ui, and MDMS repositories.
Work
Product ownership backed by delivery evidence.
Selected systems and programmes across enterprise workflows, search, document intelligence, release infrastructure, and nationwide public services.
- 70%
- faster search
- ~1.9m
- 352-test suite
- 200+
- merged contributions
Case study · Document intelligence and search
Building document intelligence and production search for enterprise workflows
What shipped
On-demand text extraction that made PDFs, documents, and images viewable without downloading and searchable alongside workflow data.
problem before AI
- +Users downloaded and opened every attachment to understand its contents.
- +Rich-text content was searchable, but information inside PDF, DOC, and image attachments was not.
- +Reviewers spent significant time extracting relevant information.
- +Important attachment context was unavailable to search and reviewers.
On-demand AI extraction was added so users could inspect those files inside the workflow without downloading them, while also turning their contents into searchable context for search and review.
document intelligence
I built upload validation and preprocessing around files added through the MDXEditor-based rich-text editor. For supported PDF, DOC, and image attachments, users trigger the Anthropic-powered pipeline through an “Extract text” action, then view the extracted content without downloading the file. I also extended the global search capability I had previously built so that extracted content was searchable alongside workflow records.
Technical leadership
Led architecture decisions for the parser and AST model, PostgreSQL and GIN indexing, user-triggered extraction, and the team implementation approach.
attachment-to-search pipeline
MDXEditor
→PDF, DOC, or image upload
validation
→allow supported types
Extract text
→user starts extraction
preprocessing
→prepare the attachment
Anthropic SDK
→extract document context
view + search
show and index the text
MDXEditor provided the rich-text editing surface; I implemented the product integration and the user-triggered processing, extraction, viewing, and search-indexing flow around it. Uploading a file alone does not run extraction.
Unsupported file types are not allowed to upload. Validation blocks the upload action and shows a toast before any file reaches the extraction pipeline.
Before
Search already covered rich-text workflow content, but it could not see information inside uploaded PDFs, DOC files, or images. Users had to download those attachments, open them separately, and inspect them manually.
After
Users can select “Extract text” for a supported attachment, inspect its content inside the workflow without downloading it, and find that content through the existing global search.
also part of this platform
- +Lucene-style query syntax for field-specific and free-text search
- +Validated PDF, DOC, and image uploads as workflow attachments
- +Prisma-backed indexing after profiling flagged large-dataset queries as the SSR bottleneck
- +SSR rendering optimized for pages under heavy data load
Case study · Engineering productivity
Designing a multi-agent implementation review pipeline
Why it exists
Static analysis catches rule violations, not implementation judgment. I designed a staged workflow that scopes a change, implements and tests it, then reviews the diff before it reaches a human reviewer.
plan agentscopes the change
→implement agentcodes + runs tests
→review agentseverity findings
→human reviewevidence before merge
high
blocking — logic errors, security
medium
should fix — edge cases, perf
low
nice to have — style, naming
The agents produce implementation and review evidence; a human remains the decision-maker before merge.
Case study · Quality and release infrastructure
Scaling E2E testing without slowing the team down
Outcome
A deterministic 352-test Playwright suite running in about 1.9 minutes across 8 workers, with retries reduced from two to zero.
problem
As the workflow-management platform grew, our Playwright suite grew with it. Running it serially meant a full regression pass ate into every merge — slow feedback loops make people skip tests, which is worse than not having them.
approach
Set up a GitLab CI/CD pipeline that shards the suite across 8 parallel Chrome workers inside Docker, so local and CI environments stay identical. Verified the sharding strategy scales cleanly beyond 8 workers before committing to that count, with multi-browser runs planned next.
Technical leadership
Defined the Playwright, Vitest, seeded-database, Docker, and parallel CI strategy, aligning the team on a repeatable approach for extending coverage.
pipeline
commit / MRGitLab
→Docker buildconsistent env
→8 parallel workersChrome
→352 tests~1.9 min
→mergevalidated
reliability, not just speed
Parallelizing the suite fixed runtime, but the earlier version also had flaky failures that needed retries to pass reliably in CI. Root-caused the flakiness and got the suite down to zero retries and zero failures.
before
retries: 2
failures: flaky, recurring
after
retries: 0
failures: 0
integration tests, not mocks
Migrated the suite from Jest to Vitest — but more importantly, these are Vitest-based integration tests that run function-level business logic against a real, seeded test database rather than a mocked DB layer. That catches a category of bugs (query behavior, schema drift, join issues) that mocks structurally can't, and it's a large part of why the suite's reliability improved alongside its speed.
352
E2E tests
8×
parallel workers
~1.9m
full suite runtime
2 → 0
retries needed
Case study · Digital public infrastructure
Shipping across DIGIT's public-service delivery stack.
At Refixd Technologies, I contributed to eGovernment Foundation's open-source DIGIT ecosystem across sanitation, public health, rural water, shared frontend infrastructure, platform configuration, and production delivery.
- 198
- merged PRs in the public GitHub record
- 10
- DIGIT repositories contributed to
- 56
- merged PRs in SANITATION
- 30
- merged PRs in the health field app
Technical delivery details
Public-health delivery
Built and refined household and campaign workflows in the Flutter field-worker app, then connected frontend, role, localisation, dashboard, persister, indexer, and environment configuration for releases.
Rural water operations
Contributed employee workflows, role configuration, payment-related journeys, validation, search, and release fixes to Punjab mGramSeva, a DIGIT-based rural water-management implementation.
Shared platform and releases
Worked across shared frontend libraries, responsive typography, master data, Helm values, deployment configuration, dashboards, reports, indexers, and persisters rather than treating the UI as an isolated layer.
How the work crossed platform boundaries
DIGIT core
Reusable workflows and platform services
Mission modules
Sanitation, health, and rural water
Delivery surfaces
Citizen web, employee web, and field apps
Operations
MDMS, dashboards, DevOps, and releases
# projects
Jal Jeevan Mission
Jan — Jul 2024eGovernment Foundation
Delivered React and Flutter modules for a national rural water-access mission; integrated the SBI-ePay payment gateway for Madhya Pradesh.
ReactFlutterPayments
mGramseva Punjab
Jan — Jul 2024eGovernment Foundation
Shipped to Android and iOS; extended functionality for metered and non-metered water connections and an HRMS module.
React NativeMobile
Frontline Worker's App
Jul 2023 — Jan 2024eGovernment Foundation — Health Campaign
Built a mobile app for field health workers to find eligible beneficiaries, capture medicine side-effects, and track resource inflow/outflow across regions.
FlutterHealthcare
Digit Sanitation
Feb 2022 — Jun 2023eGovernment Foundation
Strengthened a shared React component library (Storybook, npm-published, WCAG-compliant) used by Urban Local Bodies nationwide; mentored NIC and PwC engineers.
ReactStorybooka11yAWS S3
Pivotate
Aug 2020 — Mar 2021Pivotate
Led UI design and implementation on a proprietary framework; built backend services and REST/GraphQL APIs (Apollo Server) with Node.js and Express.js, using React Redux and React Query to manage complex client-side state.
ReactApollo GraphQLReact ReduxReact QueryNode.js