Map what matters.
- Interview leaders and domain owners.
- Walk code and architecture with engineers.
- Read telemetry, incidents, and release history.
Modernize critical systems. Without betting the business.
Eight verified efforts show the same operating model: protect continuity, make change reversible, and prove contracts before changing production. In recent work, AI accelerated analysis, implementation, and verification while humans retained architecture, sequencing, and release accountability.
- Almost $1Mannual infrastructure savings
- $80Kannual workflow savings
- 4-weekAI-assisted backend rebuild
- 10x growthsupported as accuracy reached 99.99%
Portfolio scope: This portfolio combines work delivered through Unravel with modernization experience gained in prior operating roles.
Operating model
Judgment before machinery.
Bound business risk.
Define continuity, customer, revenue, and data constraints before choosing target architecture.
Proof: AWS continuity, OneStudyTeam workflow sequencing, YUI coexistence.
Map contracts and dependencies.
Identify consumers, readers, writers, wire formats, and invariants that must survive.
Proof: Celerity mobile contracts, OneStudyTeam data contracts.
Create reversible paths.
Use coexistence, dual operation, staged cohorts, and versioned opt-in instead of one-way cutovers.
Proof: YugabyteDB dual paths, AWS cohorts, shared SDK opt-in.
Accelerate bounded work with AI.
Let AI analyze, implement, and test only after humans define architecture, acceptance gates, and release safety.
Proof: Celerity and OneStudyTeam.
AI changes throughput. Not accountability.
Recent work uses AI where the modernization boundary is explicit and machine-speed output can be checked against real behavior.
- Humans own
- Architecture, contracts, sequencing, stakeholder trade-offs, continuity, and release safety.
- AI accelerates
- Codebase analysis, implementation, synthetic data, property-based tests, and verification work.
- Evidence gates
- Celerity used 101 client contract assertions and 247 live-stack checks. OneStudyTeam used synthetic data and property-based tests to fuzz data combinations.
Three cases show speed, business value, and production continuity.
Start with an AI-assisted backend rebuild, a customer-workflow replacement, and a zero-downtime cloud modernization. Expand the timeline for broader experience across frontend, data, databases, and SDKs.
-
2026
Celerity
Java → Python backend Returned direct product change control and reduced maintenance surface by 55%.Unreleased target Business impactRemoved vendor lock-in, returned direct product change control, reduced source maintenance surface 55%, and created 4 independently deployable and scalable processes.
Duration4 active weeks. Original Java build took about 21 active weeks. Two more weeks would have avoided overtime.
ComplexityHigh. Two strict mobile contracts, Kafka compatibility, hidden framework behavior, data and storage invariants, and a compressed finish.
Why it was hard
Replacing a Java Quarkus monolith with four Python FastAPI processes had to preserve live iOS, Android, Kafka, data, and storage contracts. The target remained unreleased, so this was not a production cutover.
How risk was controlled
- Extract contracts from mobile clients.
- Verify API, database, and storage behavior against the real stack.
- Run 101 client contract assertions and 247 live-stack checks at each gate.
AI-assisted?Yes. Claude Code agents did substantial build and verification work; humans defined contracts, gates, and architectural boundaries.
ReferenceAjay Madhok, CEO, Celerity Studio
Read Terra Incognita, the detailed modernization account → -
2025
OneStudyTeam
Retool → React + Clojure $80K annual savings, with no downtime, no outage, and no loss of customizations.B2B healthcare SaaS Business impact$80K saved per year, zero downtime, and no outage with no loss of customizations. Customer UX improved while the modernized stack stayed maintainable by the engineering team.
DurationApril through June 2025, under 3 months elapsed. 2 senior engineers focused on AI-assisted modernization delivery.
ComplexityMedium. Integrations and data contracts, relational schema with 2,000 columns, production data movement, and multi-team/stakeholder coordination. Warehouse redesign was included for performance under human judgment.
Why it was hard
OneStudyTeam depended on an unmaintained Retool system with more than 30,000 lines, frequent customization loss, falling customer NPS, a 2,000-column relational schema, and multiple teams and owners.
How risk was controlled
- Define boundaries with users and owners, then move one workflow at a time.
- Scope backend, frontend, data, and warehouse performance together.
- Test, gather feedback, iterate, ship, then retire each old workflow.
AI-assisted?Yes. AI analyzed the 2,000-column relational schema and existing codebase, then generated synthetic data and property-based tests to fuzz data combinations.
Human judgment2 senior engineers owned workflow sequencing, integration and data-contract boundaries, continuity gates, cross-team decisions, stakeholder alignment, and a warehouse redesign to improve performance.
ReferenceRaman Adlakha, India Head and Director of Engineering, OneStudyTeam
-
2022
Helpshift
Hadoop + Hive + HBase → Spark + Redshift Customer NPS rose from 4 to 25 because data became more accurate and refresh time fell from 24 hours to 1.5 hours.Zero downtime Business impactCustomer NPS rose from 4 to 25 because data became more accurate and refreshed faster: accuracy improved from 90% to 99.99% in one year, and refresh time fell from 24 hours to 1.5 hours. The platform supported 10x growth while strengthening resiliency, scalability, and adoption. Helpshift's Chief Customer Success Officer tracked NPS.
Duration14 months starting October 2021. The team had 4 people dedicated to the work: 2 senior and 2 junior. The one-year accuracy period is an outcome window, not modernization duration.
ComplexityHigh. Old and new systems stayed live and continuously synced during a 16-pipeline transition spanning 2 billion devices and 1 PB daily ingestion. Customer retention and billing metrics could not pause or break; this was customer billing data, not Helpshift billing.
Why it was hard
Hortonworks stopped supporting Hadoop, HBase, and Hive. Recurring HBase and Hive outages often forced teams to rebuild HBase data, while customer retention and billing metrics could not pause or break.
How risk was controlled
Both systems stayed live and continuously synced. The team rewrote two critical pipelines on Spark and Redshift, supported by Hudi, PostgreSQL, Arrow, Airflow, and DBT, then ran old and new paths in parallel with data quality checks. Once the new outputs were trusted, the team shut down the old paths for those two pipelines and repeated the pattern for 14 more pipelines. The old system was then shut down with zero downtime.
Delivery modelHuman-led; no AI-assisted coding.
Human judgmentSenior architect judgment governed contracts, architecture, sequencing, gates, tradeoffs, and risk calls.
Interview evidenceKiran Kulkarni. ReferenceAkshay Vaidya, Head of Engineering, Helpshift
-
2022
Helpshift
MongoDB → YugabyteDB More consistent data operations, greater scalability, and headroom for future load.Zero downtime Business impactMore consistent data operations, greater scalability, and headroom for future load without interrupting customer service.
DurationAbout 6 months elapsed.
ComplexityHigh. Live 2 TB store, data integrity, dual reads and writes, and a 20-plus phase roadmap.
Why it was hard
A live 2 TB MongoDB primary store had to move to YugabyteDB distributed SQL, with supporting service extraction, without compromising data integrity or customer availability.
How risk was controlled
- Build a staging environment.
- Run both databases through dual reads and writes.
- Move gradually across more than 20 phases.
Delivery modelHuman-led; no AI-assisted coding.
-
2021
2022Helpshift
Native SDKs → shared JS core Faster SDK delivery, consistent behavior across platforms, and lower duplicated maintenance cost.Customer-paced rollout Business impactFaster SDK delivery, consistent behavior across platforms, lower duplicated maintenance cost, and broader customer upgrade reach.
Duration4 to 6 months elapsed.
ComplexityHigh. Multi-platform compatibility, version support, customer-controlled upgrades, and coordinated releases.
Why it was hard
Separate iOS, Android, and Unity or game-engine SDKs had to converge on a shared JavaScript core while existing native versions remained supported on customer-controlled timelines.
How risk was controlled
Validate shared behavior across all three platform families, then release through versioned opt-in so customers controlled upgrade timing.
Delivery modelHuman-led; no AI-assisted coding.
-
2021
Helpshift
AWS California → Virginia Almost $1M annual infrastructure savings, completed with zero downtime and no data corruption.Zero downtime Business impactAlmost $1M in annual infrastructure savings. Production cutover completed with zero downtime and no data corruption, while establishing a more current operating baseline.
Duration9 months, January through October 2021, from active planning through production rollout. California decommissioning continued afterward.
ComplexityHigh. About 1,000 instances, 120 application services, 60K requests per second, 1 PB of S3 data, 150 TB of Kylin data, and multiple state and traffic strategies.
Why it was hard
State, applications, traffic, release, monitoring, and infrastructure baselines had to move between AWS regions for a B2B portfolio spanning a long tail and very large customers.
How risk was controlled
- Map service dependencies, readers, writers, and data entities.
- Establish replicated event and state paths.
- Move traffic by customer cohort, from test domains to top accounts.
Delivery modelHuman-led; no AI-assisted coding.
ReferenceAkshay Vaidya, Head of Engineering, Helpshift
Read the full AWS modernization deep dive → -
2015
2017Helpshift
YUI → React Product roadmap never paused while React improved future delivery and stability.Live system Business impactProduct roadmap never paused. React improved future delivery, performance, reliability, consistency, usability, and extensibility.
DurationAbout 2 years elapsed.
ComplexityHigh. Live product, broad UI surface, long coexistence, and uninterrupted feature work.
Why it was hard
A broad, live YUI frontend had to become React while product teams continued shipping features without pausing the roadmap.
How risk was controlled
Move incrementally by product surface while YUI and React coexist, preserving delivery throughout the two-year transition.
Delivery modelHuman-led; pre-LLM delivery.
Modernization phase zero
Three weeks to find the right first move.
Every case above started by finding what could not break. Discovery maps those constraints, tests the highest-risk seams, and produces an executable phase one before architecture hardens into commitment.
Test the weak seams.
- Choose spikes, operational shadowing, or data analysis per risk.
- Verify contracts, rollback behavior, and data correctness.
- Separate architectural symptoms from business causes.
Make phase one executable.
- Deliver Prioritized risk map.
- Deliver Detailed phase-one plan.
- Deliver Directional roadmap.
Decision at week three: Approve phase one.
Decision package: known constraints, tested risks, explicit acceptance gates, and an executable phase-one plan.
Sequence: diagnose correctly, modernize safely, then use AI where verification makes acceleration responsible.