Bengaluru · India · @dgtalbug

Balasubramaniyan Murugappa

Senior Software Engineer,
building distributed systems that scale taking AI/ML models to production streaming events at burst scale building tools developers love

9+ years designing and operating distributed, cloud-native, event-driven, and AI/ML systems across GCP and AWS — owned from architecture and ingestion through deployment, observability, and production support.

Explore the work Resume

// stack running in production

// 01 · production impact

Recent work, observed in production.

Nine years of building production systems — payments, network automation, enterprise analytics, retail intelligence. Here are the two most recent, shown end to end: signals enter, intelligence is produced, services respond, and operations can see what happened.

/01

Shipping intelligence

Carrier and shipment events become features, model inputs, online predictions, and operational delivery windows.

Why it mattersHelps operations and customers know when a shipment will actually arrive — not a raw model score.
Production pathstreaming carriereventsIngest · Beam / WebFluxFeatures · LightGBMCloud Run serving · 140 RPSETA window · 7–20 min
  1. Carrier events
  2. Beam / WebFlux
  3. Feature workflow
  4. Model serving
  5. Delivery window

Owned end to end: ingestion, feature engineering, model workflow, serving, observability, and production support.

3M+ records/day
Event processing

Apache Beam and Spring WebFlux process carrier and shipment events for downstream prediction workflows.

140 RPS
Online inference

Cloud Run prediction serving sustains production traffic at ~80 ms p50 latency.

7–20 min
ETA window

Out-for-delivery estimates use practical windows; final delivery is approximately ±2 hours.

/02

Network fault intelligence

Bursty SNMP events become normalized, classified signals that support operations and self-healing workflows.

Why it mattersRaw telecom network traps become actionable faults fast enough for near-real-time operations.
Production pathingesting ~9,600 traps/seccircularbufferClassify · LightGBMself-healedops action
  1. SNMP / UDP
  2. Circular buffer
  3. WebFlux
  4. Kafka
  5. Fault action

Back-pressure, failure handling, observability, and operational support are part of the design — not an afterthought taped on during an incident.

~9,600 traps/sec
Real-time ingestion

UDP ingestion and circular buffering absorb bursty network traffic before reactive processing.

~20% fewer faults
Self-healing

LightGBM-assisted analysis powers classification and self-healing paths for recurring faults.

24×7 operations
Operated, not demoed

Instrumentation, recovery paths, and production support were included in the system boundary.

// 02 · selected work

Systems built to explain themselves.

Four open engineering projects connecting a problem, a system shape, and the choices behind it.

/01

Dextree

Local-first code intelligence

Parses repositories into semantic graphs for navigation, impact analysis, and Git-aware blast-radius checks.

commitimpactedchange → semantic graph → blast radius
  1. IDE / CLI
  2. Tree-sitter + LSP
  3. DuckDB / DuckPGQ
  4. Impact engine
Engineering thesis

Understand change impact before shipping it, while keeping source code local.

View repository
/02

ARC Framework

Cloud-native agent runtime · co-developer

A runtime for building and operating agent workflows with explicit orchestration, contracts, attribution, and telemetry.

ARC runtimeagent · planagent · codeagent · evaltools · MCPorchestrate · observecontracts in → attributed work out
  1. Clients
  2. API contracts
  3. Agent runtime
  4. Workers + tools
  5. Observability
Engineering thesis

Agents need operational boundaries: back-pressure, state, identity, policy, and evidence.

View organisation
/03

Guardrails

AI inspection + policy enforcement

Inspects AI requests and responses for PII, prompt injection, toxicity, and versioned policy decisions.

audit trailrequestsallowedPII · injection · toxicityallow / redact / block — every decision versioned
  1. Request
  2. Detectors
  3. Policy engine
  4. Allow / block / redact
  5. Audit trail
Engineering thesis

Safety controls should be observable, versioned, and explainable to the operator.

View repository
/04

Cerebro

Model reliability + introspection

Versions model artifacts, validates contracts, explains predictions, and monitors quality and drift.

SHAP · why this predictioncarrierdistancehourroutedrift monitor · livethresholdversioned artifacts · alert before it liesv7 · reproducible
  1. Data + features
  2. LightGBM
  3. Model registry
  4. FastAPI serving
  5. SHAP + drift
Engineering thesis

Prediction confidence comes from reproducibility, explainability, and visible operational guardrails.

View repository

// 03 · engineering surface

From signals to operated intelligence.

An architectural map, not a badge wall — each layer says what the technology is responsible for.

01 · data & events

Move signals safely

Delivery semantics, buffering, throughput, and replay behavior.

Kafkaevent streaming
Apache Pulsarstreaming + queues
NATSlightweight messaging
Pub/SubGCP messaging
Apache Beampipeline execution
02 · state & analytics

Keep state trustworthy

Durable stores, warehouses, graphs, and caches.

PostgreSQLdurable state + pgvector
BigQuerywarehouse analytics
Neo4jgraph store
Rediscache + queues
DuckDBlocal analytics
03 · services & contracts

Build clear contracts

Typed services, reactive runtimes, and tooling.

Javaproduction services
SpringWebFlux + Boot
PythonML + automation
Goruntimes + tooling
TypeScriptReact + Node
04 · models & ML

Version model behavior

Training, explainability, registries, and serving.

LightGBMprediction
SHAPexplainability
MLflowartifact registry
Notebooksexploration + evals
FastAPImodel APIs
05 · AI platform & agents

Operate agents responsibly

Managed ML, orchestration, protocols, and policy.

Vertex AImanaged ML + LLMs
Agent GardenGCP agent platform
LangGraphagent orchestration
MCPtool protocol
Guardrailsinspection + policy
06 · runtime & operations

Operate what ships

Cloud runtime, containers, traces, and recovery.

Cloud Runserverless serving
GCSobject storage
Kubernetesorchestration
Dockerdelivery units
ECS · EKSAWS runtime
OpenTelemetrytraces + metrics

// 04 · experience

Nine years across systems, platforms, and products.

From full-stack products to platform engineering, high-throughput event systems, and AI/ML-enabled production services.

  1. Jun 2025 — Present

    Best Buy India

    Senior Software Engineer

    Shipping intelligence across ingestion, model workflows, serving, observability, and support.

  2. May 2024 — Jun 2025

    Tejas Networks

    Senior Software Engineer

    High-throughput network fault management and LightGBM-assisted self-healing.

  3. Oct 2021 — May 2024

    Synechron (Attra)

    Senior Associate Technology

    Payments, open banking, AWS delivery automation, and AI-assisted internal workflows.

  4. Sep 2020 — Oct 2021

    HCL Technologies

    Senior Engineer

    Cisco network automation, production debugging, and root-cause analysis.

  5. May 2019 — Aug 2020

    Visual BI Solutions (Atos)

    Developer

    Reusable visualization components and Node.js services for enterprise analytics.

  6. May 2017 — Apr 2019

    Trikore Software Solutions

    Software Developer

    Web and Android products across CRM, logistics, and education.

// 05 · engineering principles

Operational maturity is part of the architecture.

The useful design is the one that still makes sense when the queue is backed up, the model drifts, or an operator needs an answer at 2 a.m.

01

Clarity before cleverness

Explicit contracts and readable failure modes beat abstraction for its own sake.

02

Own the complete path

Ingestion, serving, observability, support, and recovery belong inside the system boundary.

03

Data informs the API

Reliable signals produce reliable features, models, services, and decisions.

04

Automate repetitive work

Tools should amplify engineering judgment, not merely relocate toil.

05

Measure before tuning

Instrument the real path, identify the constraint, then optimize with evidence.

06

Earn the complexity

Add infrastructure only when simpler boundaries cannot satisfy the requirement.

// 06 · education & recognition

Foundations, formally earned.

M.Tech · 2026

BITS Pilani (WILP)

M.Tech in Cloud Computing. Dissertation: design and implementation of a privacy-first data guardrail engine.

PG Certificate · in progress, 2026

IIT Roorkee

Postgraduate Certificate in Forward Deployed Engineering.

B.Tech · 2017

Anna University

B.Tech in Information Technology, Arunai Engineering College.

  • AWS Certified Cloud Practitioner
  • Winner · Google Android Developer Scholarship, 2018

// contact

Build systems that hold up in production.

Distributed systems, AI/ML platforms, forward-deployed engineering, and developer infrastructure. The resume has the details; GitHub has the proof.

View resume GitHub LinkedIn