Portfolio

Selected Work

Case studies from enterprise AI, regulated FinTech, and high-scale commerce. Each one includes what I assumed going in, what the reality turned out to be, and what I learned.

Case Study 01 · Deloitte · 2022–2026

GenAI Copilot for Tier-1 Enterprise FinTech

35+ tier-1 clients · 200% platform engagement growth · 10–15x retrieval velocity · AWS Lambda, OpenAI API, RAG, KYC/AML architecture

What I assumed walking in

The client request seemed straightforward — they wanted better AI tooling and smarter prompting layers. My initial assumption was that this was a capability gap. I thought we needed to build better AI features on top of an existing data layer.

What the reality turned out to be

The moment I audited the backend, I found the real problem: catastrophic data fragmentation across 9 legacy enterprise systems. The intelligence required to make an LLM functional was siloed across disconnected applications with conflicting schemas. Building AI on top of that would have produced expensive hallucinations and broken compliance contracts.

The moment that confirmed it

During an early engineering sync, an alpha model attempted to execute an agentic workflow using un-unified data streams. It pulled conflicting customer profiles and produced a hallucinated, non-compliant output. In a regulated environment, that is a zero-tolerance failure. That near-miss gave me the leverage to halt premature rollout and force the foundational rebuild first.

What I got wrong mid-way

I initially over-indexed on optimising RAG pipeline chunking strategies, assuming we could fix output quality at the retrieval layer. After two sprints of flat precision metrics, I realised I was patching a structural data problem with fine-tuning. I paused vector database tuning and redesigned around a centralised Knowledge Graph — the decision that eventually unlocked 10–15x execution velocity.

What I learned

The most important PM skill in regulated AI is knowing what not to ship and when. Trust is not a launch checklist — it is the product itself. And when precision metrics do not move after two sprints of retrieval-layer optimisation, the problem is almost always structural, not parametric.

35+Tier-1 clients
200%Engagement growth
10–15xRetrieval velocity
20%→70%Service efficiency
Case Study 02 · TCS · 2020–2021

Retail Wealth and Pension Platform — Major European Financial Group

Consumer-facing retail investing · Pension management · Monte Carlo simulations · Automated risk profiling · Best Performance of Year award

What I assumed walking in

Coming from an engineering background, my bias was that more control equals more value. I fully believed in the manual asset-rebalancing dashboard — granular multi-variable toggles, data-dense interfaces. I assumed retail investors wanted a powerful manual trading-desk experience.

What user research actually showed

Two weeks into active development, I watched users interact with our Figma prototypes. Instead of feeling empowered, they froze. They were anxious about making a mistake with their actual pensions. User after user asked the same question: can the platform just do this for me? The drop-off intent data was unambiguous.

The kill

I stopped the manual rebalancing dashboard mid-sprint. The hardest part was overcoming months of signed-off business requirements. I brought the raw usability session recordings and persona maps into the room and showed stakeholders what was actually happening. I redirected engineering to automated Monte Carlo retirement simulations — users input retirement age and risk tolerance, backend handles the math.

What I learned

The most dangerous assumption in consumer FinTech is that more control equals more value. Pension users are not traders — they want peace of mind. I now run a cognitive load audit before any consumer-facing feature goes to scoping: does the user actually need to make this decision, or can the system make it better?

3xUser adoption
100%Manual steps eliminated
Best Performerof Year, TCS
Case Study 03 · Shiprocket · 2021–2022

Seller Intelligence Platform — 100,000+ SME Merchants

100k+ active sellers · 17+ carrier partners · Event-driven resolution system · 85% operational efficiency gain

What I assumed walking in

My immediate assumption was that non-delivery failures were a supply-side and carrier execution problem — unoptimised routing algorithms, missed addresses, dropped mapping API connections. I expected to spend most of my time on carrier-side tooling.

What cohort analysis actually revealed

Merchant cohort data showed something completely different. Merchants were purely reactive — they would not check delivery anomalies until days after a package had already failed and been routed back as a total loss. This was not a carrier execution problem. It was an information latency problem.

What I built

An event-driven automated resolution system. The moment a carrier logged an exception on the ground, the platform fired a real-time WhatsApp/SMS trigger directly to the end consumer to verify availability or update address details. Resolution loop automated at the moment of friction — no merchant bottleneck, no lost packages.

What I learned

At hypergrowth velocity, the instinct is to look at the visibly failing system and fix it directly. In this case the visible failure was carrier routing. The real failure was upstream in the information flow. I now start every operational problem investigation by mapping the full information lifecycle, not just the execution layer.

100k+Active sellers
85%Ops efficiency gain
17+Carrier partners
18%Overhead reduced

PM Judgment — The Calls That Defined the Outcomes

Three moments where the failure was invisible until I looked for it.

Hidden Failure

The dashboards were green. Retention was dying.

D7/D30 cohort analysis in Amplitude and SQL revealed retention collapsing while executive dashboards showed healthy CTR. We were reaching users across push, SMS, and email with zero coordination. Nobody flagged it because the surface metrics were green. I re-architected the notification engine with event-driven triggers and frequency capping — driving 200% sustainable engagement uplift.

Production Incident Averted

Engineering missed it. I caught it in the PRD.

While writing acceptance criteria for a wealth management API pipeline, I identified synchronous KYC/AML API calls that would have cascaded to server failure during high-volatility trading. Engineering had assumed synchronous responses. I mandated async queue architecture and graceful degradation states before a line of code was written. The incident never happened.

Scope Won

I blocked marketing scripts from the transactional pipeline.

Marketing stakeholders had already convinced team leads to hardcode banner campaigns into the AWS CDK transactional infrastructure — it was in sprint planning as an inflight task. I stopped it with a RICE/Kano trade-off map and built a decoupled Braze/Segment self-serve infrastructure instead. Marketing launch time dropped from weeks to days. Platform velocity preserved.

Recognition

🏆

Top Impactor Award

Deloitte — platform impact across 35+ enterprise FinTech clients

Spot Award

Deloitte — zero-disruption data migration delivered ahead of schedule

🥇

Best Performance of Year

TCS — cloud modernisation and wealth platform delivery

💻

Google Summer of Code

UNESCO & Ushahidi · 2018 & 2019 — selected from under 8% acceptance rate