Janahan Thavendran

RAG sales assistant

Shipped a multi-agent assistant, designed from the questions reps actually asked, that briefs 700+ sales, customer success and account managers before, during and after calls. Hybrid keyword and vector retrieval with reranking across product docs, CRM notes and call transcripts keeps similar plan and feature names apart.

Commercial impact

Lead-to-close conversion doubled after rollout with no increase in lead volume. Call preparation and follow-up time fell, freeing reps for higher-intent prospects, and customer satisfaction rose.

Evals before launch

Defined the evaluation strategy: a 300-question gold set built from real sales and customer-success questions with human-reviewed answers, scoring retrieval, correctness, faithfulness and correct refusal through an LLM judge spot-checked by sales reviewers. A silver set from live interactions and rep-flagged errors keeps it growing, and it runs on every prompt, retrieval, model and content change.

Model choice

Selected models against the eval set: Claude 3.5 Sonnet for multi-step reasoning and synthesis, Claude 3.5 Haiku for lighter, high-volume tasks, holding answer quality while controlling cost and latency.

Multi-agent architecture

Specialist agents query databases and internal services through tool calls and MCP alongside the RAG layer. All agent access is read-only, so the assistant briefs reps on live account data without being able to change it.

Trust by design

Every answer shows its source, owner, last update and a confidence level. Pricing is confirmed with its owner, conflicting sources are flagged, and with no reliable source the assistant says it doesn't know and drafts a question to the document owner. Access follows each user's permissions, with tracing on every call.

Refunds digital worker

Shipped an AI digital worker for refunds and chargebacks: the model reads the case, while policy rules in code, idempotent ledger payouts and human approval for exceptions decide the money, with every action logged to the CRM.

Embedded finance

Turned payments into a revenue line for SaaS platforms: sub-merchant onboarding under PayFac models and unified multi-PSP integrations (Stripe, Adyen, Airwallex and others) that simplified routing, settlement and reconciliation.

Platform strategy

For an EdTech SaaS client, reframed a utility product as a platform: a new payment layer, a marketplace model, a product for solo instructors, and Year 1 acquisition targets across new customer segments.

Team and P&L

Led 10 direct reports (6 PMs, 2 designers, 2 analysts) in a 25+ person org; ran OKRs and promotions. Co-owned the payments P&L across 10+ markets, including buy and sell rates, pricing and staffing. Payments revenue grew 28%.

PayFac strategy and launch

Built the business case to move from merchant of record to a PayFac model, capturing payments margin and owning onboarding, payouts and reconciliation. Reached $10M monthly GMV in two months and $50M in six. Designed idempotent, event-driven transaction flows and treated reconciliation as a product: unreconciled items fell 85% and settlement close was 70% faster.

ML fraud scoring

Shipped fraud scoring on unattended payment flows, combining in-house models with PSP risk tools, threshold tuning and a manual review queue for borderline cases. Fraud losses fell 30%.

Payment routing

Led routing strategy across 20+ payment rails (Adyen, Stripe, Checkout.com, local acquirers) and 25+ payment methods, lifting authorisation 2 pp and cutting processing cost per transaction 9%.

API platform

Unified a post-acquisition API estate into one platform with OpenAPI contracts, versioning, SDKs, a developer portal and shared event streams. API adoption tripled, developer NPS rose 40 points, and partner onboarding went from weeks to days.

Dispute automation

Automated chargeback evidence assembly from order logs, delivery proof and PSP webhooks, lifting dispute recovery 25%.

Recommendations at scale

Owned the recommendation product through the TabSquare acquisition and scaled it to 12 markets, keeping holdout tests, eval gates and latency budgets in place.

ML to production

Took recommendation models from experiment to production across 8,000 merchants. Real-time personalisation lifted average order value 8% against a live holdout, with no added checkout time.

Problem framing

Turned a vague "use AI to grow revenue" mandate into one measurable problem: acceptance rate as the driver metric, with checkout time, completion rate and latency as guardrails.

Data before models

Designed a 25+ event tracking plan across four channels before any model was trained. Each recommendation was tagged model-driven or rule-based and linked to its API response, list position and anchor item; clicks, adds, removals and group size were joined to transactions for revenue ground truth. This made ML-vs-baseline comparison and position-bias correction possible.

Cold start, drift and serving

Handled anonymous sessions, new items and changing menus with shared item attributes, recency weighting, live availability filtering and fallbacks. Set a model-size cap and cost per request before rollout to keep latency under a second.

Partner platform

Owned the middleware syncing catalog, orders, loyalty and payments with 50+ partners, including Google Pay Singapore and the OCBC Bank app, which showed merchant menus and sent orders straight to merchants.

Core platform

Owned the commerce platform shared by every channel, with architecture decisions on distributed services, real-time transactions, PostgreSQL read replicas, Redis caching and observability. Led 10+ engineers in a 20+ person team; during this time the company grew merchants 10x and revenue 4x and raised a $10M Series B.