Dwarka Prasad
Bairwa

• Led an Infrastructure-as-Code migration of 8 mixed-stack projects (Spring Boot, Node.js, Python, Flutter) from Docker Swarm to Kubernetes using Terraform, Helm, and versioned manifests — with Jenkins CI/CD pipelines for every service, lowering infrastructure cost 18% and infra-related incidents 80%.
• Architected a Video-KYC and video personal-discussion platform on Spring Boot microservices with an Angular frontend, serving real-time banking journeys for multiple banking clients.
• Engineered a real-time journey-monitoring system with Java REST APIs, improving reliability and reducing downtime 30%.
• Drove production release fixes and root-cause analysis during high-traffic festive seasons, coordinating client stakeholders to safeguard uptime under peak load.
• Developed a high-throughput URL-shortening service in Lua and automated recurring jobs with Kubernetes CronJobs.
• Mentor a team of 6 engineers in an Agile/Scrum environment — code reviews, architecture guidance, and cross-functional issue resolution.

• Developed and maintained a video-call mobile application in Flutter, delivering a seamless cross-platform experience.
• Integrated REST APIs in Dart and optimized MongoDB queries, improving data retrieval speed 20%.
• Implemented social-media integrations including Facebook login, and owned the release-management process.

Distillr — compression pipeline for LLM inputs

Trims what is irrelevant, re-encodes what is left in the cheapest lossless format, and accounts for every token saved. site ↗ · source ↗

• Four composable stages with a strict contract: every removal carries a path, reason and preview so an audit step can flag model answers that leaned on cut content.
• Wrote a lossless TOON encoder/decoder and an auto mode that measures TOON, compact JSON and CSV on the real tokenizer (tiktoken) and keeps the cheapest; column-wise empty-field removal keeps rows tabular.
• Benchmarked on five payload types with planted facts: 102,988 → 4,947 tokens (95.2%) at 100% needle recall; the benchmark drove three design fixes before release.
• SQLite token ledger from day one (same schema for the hosted Postgres dashboard), Typer CLI, 42 tests, CI, docs site.

Regressa — catch AI regressions before your users do

An LLM observability and prompt-regression platform: every model call logged, every prompt change versioned, output quality scored automatically, and alerts when quality drops or cost and latency spike. site ↗ · source ↗

• Designed the whole system: Fastify ingest API, BullMQ workers on Redis, TimescaleDB hypertables with continuous aggregates, a Next.js 14 dashboard with a public REST API, multi-tenant from day one (org → project → key → trace).
• Regression core: SHA-256 hashing of normalized prompt templates detects versions automatically; per-version comparison of eval score, p95 latency, error rate and cost, with a line diff between versions.
• Evals and alerting: LLM-as-judge via forced tool calls, golden-set semantic similarity, sandboxed custom scoring; threshold, percent-change, rolling-baseline anomaly and budget rules to Slack, signed webhooks and email.
• A CI regression gate (npx @regressa/node gate and a GitHub Action) that blocks merges when a prompt change regresses against production; OpenTelemetry GenAI ingestion; OIDC SSO, PII redaction, retention and an audit log.
• SDKs for Node, Python, Go and Java that wrap OpenAI and Anthropic clients in two lines; ~120 unit tests across vitest, pytest, go test and JUnit; Framer Motion dashboard; docs site on GitHub Pages.

OpticTrace — declarative API telemetry & governance gateway

Like OpenAPI, but for observability: one optic.yaml controls what API traffic reveals. A Go gateway and agent with a product site, Homebrew tap and a licensed Pro edition. site ↗ · source ↗

• Tee-point request flow that hands the client the original bytes while telemetry only ever sees redacted payloads; Prometheus metrics with request attributes as dimensions.
• Traffic-inferred OpenAPI generation, a breaking-change linter for CI, a stateful mock server, and FinOps metering per consumer.
• Inner spans for database, cache and outbound calls with governed attributes, an 8-tab dashboard, and four SDKs kept at feature parity.
• Shipped as a product: releases, Homebrew formula, offline licence model for Pro features, and a GitHub Pages product site.

Video-KYC (V-CIP) platform

RBI-compliant video customer identification for banks, serving 500–600 sessions a day across multiple banking clients: link generation → PAN/Aadhaar/DigiLocker capture → officer-led video call → auditor review. Reactive Java / Spring Boot microservices, a LiveKit media server, and an Angular portal for customers, officers, auditors, and admins.

• Migrated the media stack from Kurento to LiveKit across server and client — re-architecting call-handling workflows, STUN/TURN setup, and recording — for better scalability and call quality.
• Built network-adaptive publish quality: a connection-quality-driven controller with hysteresis and flap guards stepping video between 144p and 480p, plus configurable egress recording and per-environment bandwidth ladders.
• Ensured RBI compliance with AES encryption for customer data in transit and at rest, JWT/Spring Security auth, and full audit trails over Kafka.
• Hardened active liveness (AWS Rekognition): server-side confidence verification with configurable thresholds, camera-handoff race fixes, and real confidence scores threaded end to end.
• Led platform upgrades (Java 21 → 25, Spring Boot, Kafka), JUnit/Mockito test coverage, KYC image compression, and frontend bundle-size reduction via lazy loading.

AI agent-assist for live KYC calls

A LiveKit Agents worker (Python) that joins every Video-KYC room as a hidden participant, paired with an officer-facing assist UI in the Angular portal.

• Built end to end: streaming speech-to-text (Deepgram, with Google/AWS/Whisper switchable), per-call language selection, live captions with real-time translation, and post-call summaries.
• Added V-CIP compliance-checklist tracking, vision-based document checks, and video-risk sampling on the live stream.
• Designed the caption/event transport over LiveKit data channels and the officer UI: floating assist pane, live transcript, closed-caption toggle, consent-to-transcript bridging.

API observability platform

In-house observability for API traffic and customer journeys: dual-language client SDKs emitting schema-validated events to Kafka, ingested through Telegraf into ClickHouse and visualized in real-time Grafana dashboards.

• Migrated the time-series store from InfluxDB to ClickHouse, sustaining ingestion of ~100k events/sec with low-latency dashboards — cluster topology, sharding, migrations, and ClickHouse↔S3 export/import.
• Re-engineered the Kafka ingestion SDK, cutting data-ingestion latency from 74ms to ~20ms (≈73%).
• Created an integration SDK (Java + Node/TypeScript) that simplified API instrumentation across teams, reducing integration time 40%; released to private Maven/npm registries.
• Deployed the full stack on AWS via a containerized cluster — recognized as Best Project of the Year.

Kubernetes migration & platform services

Company-wide deployment platform: roughly 1,300 Kubernetes manifests and 56 Jenkins pipelines across five environment tiers (QA → SIT → UAT → PreProd → Prod/DR) and multiple clouds — AWS EKS and bank on-prem.

• Led the Docker Swarm → Kubernetes migration of 8 mixed-stack services with Terraform, Helm, and versioned manifests: 18% lower infra cost, 80% fewer incidents.
• Owner of the engineering namespace — onboarding new services with pipelines, Vault secret integration, HPAs, and Kong routing.
• Built and operate a production URL-shortening service in OpenResty/Lua backed by Redis — no app server — with per-client custom domains, CSV-driven config reload, and IP allow-listing.
• Rebuilt Vault operational tooling as parameterized Jenkins jobs (KV ops, policy creation, rollback) with scoped token revocation, removing hardcoded credentials.

Perch — local system & developer dashboard

A self-hosted control panel for your own Linux machine: one token-protected web app on 127.0.0.1 replacing htop, df, systemctl, journalctl, docker ps, Postman, and a pile of converter tabs. site ↗ · source ↗

• Single small Python service (stdlib HTTP server, no framework) with live CPU/memory/GPU/disk/network charts, threshold alerts, process and systemd service control, storage cleanup, and package management.
• Developer toolbox — Docker containers, git repos, listening ports, HTTP tester, JSON/YAML/regex/diff converters — plus a built-in web terminal multiplexer and an AI assistant backed by the local Claude CLI.
• Shipped as a proper product: five releases to v1.5 with .deb packaging and desktop integration, GitHub Actions CI and release pipeline, a security overview tab, traffic capture, fleet view, 90-day history, and a public roadmap with issues open to contributors.

Employee Central — mobile app & backend

Internal employee platform: a Flutter app (identity, QR verification, merch store, embedded Grafana infographics, and a loan-origination demo journey) over a Node/TypeScript backend.

• Top contributor on the backend: Google OAuth hardening, employee-ID issuance, merch-store ordering, S3 document storage with presigned URLs, Postgres migrations, and the Kubernetes deployment track.
• Built the demo loan-origination journey in Flutter: camera/document capture, offer screens, and Atlas API integration.

Viveka — credit decision engine

Governance-first credit-decisioning platform: a 79-crate Rust workspace with Axum services, a sandboxed policy language, and Leptos/WASM front ends. Focused on the scoring-intake area.

• Built the merchant feature pipeline: typed feature extraction with opt-in v2 classifier weights, interpretation/EQI scoring, and taxonomy work.
• Refactored the intake server into modules and reworked intake-portal RBAC (canonical roles, maker-bench scoping, maker/checker/admin selection rules).