Forward Deployed Engineering, demonstrated
Site Reliability Engineer
Payments & Observability
Devika Tamma — Business Insights Analyst | Payments Analytics | Site Reliability Engineering
Building toward Forward Deployed Engineering — reliability and payments-domain depth, demonstrated with live systems instead of slides.
Flagship
Multi-Rail Payments Platform
A synthetic multi-channel payments processor with a live SRE observability layer on top — SLIs, SLOs, and error-budget burn computed in real time across 11 origination channels.
Simulates card, wire, ACH, and Zelle traffic across real-time and batch rails, injects randomized latency/failure incidents, and tracks per-channel and per-rail health against agreed SLAs — including a genuinely separate five-nines platform-availability metric, distinct from business approval-rate SLOs.
Live from the deployed API, right now
Platform modules
Four independent services, one shared ledger
Each of these is its own repo with its own deploy lifecycle and its own reason to exist — they read from the flagship platform's data, they don't depend on its uptime to be understood.
AI Incident Copilot
LiveWatches the platform live and diagnoses degraded channels with an LLM — without ever seeing the injected fault's ground-truth label.
Concurrent multi-incident diagnosis via asyncio.gather — fixed a real bug where sequential polling inflated later channels' reported time-to-insight by 10s+.
Reconciliation & Anomaly Detection Engine
LiveCatches the part of a payments platform a live dashboard never shows: two independent records of the same money drifting apart.
Batch-oriented and data-integrity-focused — a deliberately different discipline from Project 1's real-time service-health monitoring.
Business Insights & ROI Modeling
LiveTurns the shared ledger into the numbers a business stakeholder actually asks for: revenue by rail, a forecast with an honest accuracy score, and "what if" scenario modeling.
Payments economics modeled per rail (interchange, flat wire fees, network cost) with assumptions documented as illustrative, not hidden as fact.
Support Agent
LiveA multi-turn, guardrailed conversational agent for customer questions about a single transaction — evaluated with RAGAS and DeepEval, not a hand-rolled scorer.
RAGAS Faithfulness scored 1.00/1.00 on factual scenarios; a documented judge-model reliability finding (RAGAS vs. DeepEval HallucinationMetric) as an honest eval-selection case study.
Skills
What the platform above actually demonstrates
Reliability & Observability
- SLIs / SLOs
- Error-budget burn
- Five-nines availability modeling
- MTTD / MTTR reduction
- Incident lifecycle & on-call triage
Payments Domain
- Card (credit/debit)
- Wire (online, branch, LoanIQ, batch, IVR)
- ACH & Zelle
- Real-time + batch rails
- Reconciliation & break classification
- Interchange & payments economics
AI / LLM Systems
- LLM root-cause diagnosis
- Multi-turn guardrailed agents
- RAGAS & DeepEval evaluation
- Pluggable LLM providers (Groq, Ollama)
- Prompt & judge-model selection
Full-Stack Engineering
- FastAPI / Python
- Next.js / TypeScript
- WebSockets (live data)
- Postgres (Supabase)
- Recharts / data visualization
Delivery & Ops
- Free-tier cloud deployment
- GitHub Actions CI
- Test-driven bug discovery
- Honest trade-off documentation
- Zero-paid-infra delivery discipline
Timeline
Career, and the build log behind this platform
Oct 2025 – Present
Platform Site Reliability Engineer · Citizens Financial Group, Inc.
Payments platform reliability and observability across AWS and hybrid environments — owns end-to-end KPI/observability strategy for revenue-critical commercial banking platforms, built the EPP/MTS reporting framework from scratch (cutting MTTD by ~72 minutes on average), and reduced false-positive alerting by 30% through defined SLIs, SLOs, SLAs, and Error Budgets.
Ongoing
Hardening passes on this platform — real bugs, not just features
Fixed a sequential-polling concurrency bug, a five-nines availability-metric misclassification, and added WebSocket + timing-based regression tests — the kind of work that doesn't show up in a demo but does in a healthy system.
Jul 13–16, 2026
Expanded domain depth, shipped Reconciliation, Support Agent, and Business Insights
Grew to 11 real origination channels (added Zelle and three more wire channels), and shipped the remaining three platform modules — including an eval harness built on RAGAS and DeepEval.
Jul 10, 2026
Unified five separate demos into one platform
Recognized that presenting disconnected demo links undersold a genuinely connected system — rebuilt as one frontend with drill-down navigation into each live module.
Jul 8, 2026
Shipped AI Incident Copilot
Chose a pluggable local/free LLM strategy (Ollama, then Groq's free tier) over adding a payment method anywhere — a deliberate constraint, not a limitation.
Jul 7, 2026
Shipped the flagship Multi-Rail Observability Platform
Live SLIs/SLOs and error-budget burn across card, wire, and ACH — deployed on free-tier infra (Vercel + Render) from day one.
Sep 2024 – Oct 2025
Platform Site Reliability Engineer · Kingston Technologies, Inc. (Client: Citizens)
Automated incident detection and recovery workflows for core payment applications, scaled full-stack Datadog observability (APM, tracing, service maps, DBM/NPM correlation) across 150+ consumer and commercial applications.
Feb 2024 – Aug 2024
Cloud DevOps Engineer · Cloud Craft Tec Inc.
AWS and infrastructure-as-code work — the operational foundation the payments-reliability work above builds on.
How I build
Principles this platform was actually built under
Not a testimonials section — I don't have client quotes yet. These are real patterns from building the system above, verifiable in the commit history.
Ship live, not slides
Every project linked on this page is a real deployed service you can hit right now — including the widget above, fetching from a live API on page load.
Write evals, not vibes
AI outputs get graded against fixed scenarios with a real framework (RAGAS, DeepEval) before I trust them — including an honest writeup of where one framework's judge model was measurably less reliable than the other's.
Investigate the number that looks wrong
An "11,583x over budget" metric turned out to be a real classification bug, not a UI display issue — file rejections were being counted against a five-nines infrastructure-availability SLO they don't belong to. Absurd-looking numbers are usually real bugs.
Fix root causes, not symptoms
A sequential polling loop was quietly inflating multi-incident diagnosis time by 10s+ per extra incident. Fixed with concurrent execution and proved it with a timing-based regression test, not just a code review.
Document trade-offs, don't hide them
Free-tier hosting means cold starts. In-memory state means the backend is pinned to one instance. Both are called out directly in the relevant README as deliberate trade-offs, not glossed over.