Ajay Naidu
Software Engineer Bengaluru, India Open to hard problems

Ajay Naidu
builds for
the day
everything
breaks.

I plan for failure before launch, break it before anyone else can, and own what I ship from the first requirement to the last fix.

Now
Software Developer, PetaSight
Before
Software Developer, Morse Team
Degree
B.Tech CSE, VIT, 2025
Core
Python, TypeScript, Node.js, Next.js, PostgreSQL, AWS
See the projects Get in touch

Work

2025 – now
Apr 2026 – present · Bengaluru

PetaSightSoftware Developer

CurrentEnd-to-end ownership
  1. Owned an AI-assisted greenhouse design and quotation tool for a client end to end as the sole engineer, covering requirement discovery with their sales and engineering staff, architecture, implementation, testing, deployment and production support, and took it from prototype to an internal tool in daily use by the client’s sales team in 6 months.
  2. Designed the layout-generation and pricing engine that turns site dimensions and constraints into a complete greenhouse layout, bill of materials and priced quotation, replacing a 2-day manual design and quote process with a 30-minute design step and a quotation generated in under a minute.
  3. Built quotation revision control with full version history and diffable BOMs so sales could iterate with customers without re-running engineering calculations, cutting re-quote effort from 6 hours to under a minute.
  4. Built the ML pipeline behind a therapist “digital twin” for a healthcare platform, generating behavioural profiles from session data, running model inference, and evaluating fidelity against held-out sessions across datasets of varying scale.
  5. Ran production on AWS with Docker and GitHub Actions CI/CD, led root-cause analysis on production incidents, and shipped customer-facing releases working directly with product stakeholders.
May 2025 – Mar 2026 · Remote

Morse TeamSoftware Developer

Production
  1. Built a production digital menu, online ordering and payment platform for restaurants, owning features across the Next.js and TypeScript front end, Node.js and Express REST APIs and PostgreSQL data layer from development through live customer use.
  2. Designed the order and payment flow with explicit order state transitions, payment gateway integration, webhook verification and idempotent order creation, so retries and duplicate callbacks could not cause double charges or duplicate orders.
  3. Designed the PostgreSQL schema for menus, items, orders and payments with versioned migrations, transactional order writes and indexes on hot query paths, keeping menu search, filtering and order retrieval responsive under peak restaurant load.
  4. Built the customer facing ordering interface in Next.js with server side rendering for menu pages, route level code splitting and a reusable Tailwind component library, improving load time and keeping screens consistent across the app.
  5. Set up GitHub Actions CI to run type checks, integration tests for order and payment paths and database migrations on every pull request, catching regressions and schema issues before they reached production.

Projects

Three systems
GitHub Action · published on the Marketplace

BlastRadiusPR change-impact analyzer: what could this PR break?

TypeScriptTS Compiler APIGitHub ActionsSQLiteGit mining
  1. Designed and shipped a GitHub Action that answers “what could this PR break?” by building a symbol-level dependency graph with the TypeScript Compiler API (chosen over regex/AST parsing for cross-file and path-alias resolution accuracy) and running bounded reverse-graph BFS from changed line ranges to affected functions, endpoints, and tests.
  2. Built dependency-aware test selection that cuts CI time while preserving near-complete recall of real failures; validated by replaying historical commits and comparing selected tests against actual failures, and documented the cases static analysis cannot see (dynamic imports, DI, reflection) with fallbacks to full runs.
  3. Mined git history for churn, bug-fix density, ownership concentration, and co-change coupling to produce a calibrated 0–100 risk score, then validated it empirically against subsequent bug-fix commits rather than treating weights as assumptions.
  4. Made the Action viable on large monorepos via incremental analysis keyed by commit SHA with SQLite caching persisted in GitHub Actions cache, reducing per-PR analysis from minutes to seconds; chose in-memory adjacency lists over a graph database after benchmarking showed no need at realistic repo sizes.
  5. Published on the GitHub Marketplace, drove adoption through Show HN and direct outreach to open-source maintainers, and iterated on user-filed issues, including reducing false-positive noise after reviewers reported alert fatigue.
Durable execution engine · AI agent workflows

PhoenixTemporal-style durable workflows that survive crashes and resume at the exact step

Node.jsPython SDKPostgreSQLKEDAOpenTelemetry
  1. Designed and built a Temporal-style durable workflow engine (Node.js engine, Python SDK, PostgreSQL) using event sourcing and deterministic replay, so multi-step LLM agent workflows survive crashes and resume at the exact step within 15 s of worker death, with no re-execution of completed steps.
  2. Guaranteed effectively-once side effects by combining at-least-once execution, per-step idempotency keys, and a transactional outbox (history append + next-task enqueue in one transaction); proved it with 50 chaos runs of 500 concurrent workflows under random worker/engine kills and Postgres connection drops, with zero lost or duplicated effects verified against a ledger.
  3. Implemented the full reliability surface: SKIP LOCKED task queues, lease-based crash detection with heartbeats, three distinct timeout types, durable timers, retry policies with exponential backoff and jitter, and signals for human-in-the-loop approval, with p99 queue-to-start latency under 50 ms.
  4. Built a determinism guard that detects workflow code diverging from recorded history, plus a replay debugger and a CI gate that replays the last 200 production histories against new code in under 30 s, catching deploys that would break in-flight workflows.
  5. Ran it as a production system: stateless engine replicas, KEDA worker autoscaling on queue depth (2→20 in under 60 s), one OpenTelemetry trace per workflow with a span per activity attempt, and shared token-bucket rate limiting across workers to stay within LLM API limits; measured 12 ms/step engine overhead and 1,100 steps/sec on a single Postgres before partitioning becomes necessary.
Deep learning and computer vision

Real-time sign language recognitionLive webcam gestures rendered as text

PythonTensorFlowOpenCVCapsuleNetGRU · IndRNN
  1. Built an end to end real time system that recognises sign language gestures from live webcam video and renders them as text, running the full pipeline of frame capture, hand region extraction, preprocessing, sequence inference and text output with no offline step.
  2. Chose Capsule Networks over standard CNNs for spatial feature extraction because capsule pose vectors preserve part to whole relationships between fingers and palm, holding accuracy under rotation and viewpoint changes where the CNN baseline degraded.
  3. Modelled gesture sequences with GRU and IndRNN and compared them head to head on the same held out set, with IndRNN handling longer gesture sequences without the vanishing gradient issues the GRU variant showed.
  4. Reduced background and lighting sensitivity with ROI based hand detection, skin segmentation and per frame normalisation before inference, cutting misclassifications on cluttered backgrounds.
  5. Evaluated with a signer independent held out split so results reflect generalisation to unseen people rather than memorised hands, and documented failure cases such as similar handshapes and fast transitions along with the mitigations tried.

Skills

What I work in

Languages

  • Python
  • TypeScript
  • JavaScript
  • Java
  • Go
  • C++
  • C
  • SQL

Backend

  • Node.js
  • Express
  • REST APIs
  • gRPC
  • Protocol Buffers
  • JWT
  • OAuth
  • Microservices
  • Event Sourcing

Frontend

  • React
  • Next.js
  • HTML
  • CSS
  • Tailwind CSS
  • Material UI
  • Bootstrap

Data

  • PostgreSQL
  • MySQL
  • SQLite
  • MongoDB
  • Redis
  • Kafka
  • Alembic
  • Data Pipelines
  • ETL
  • SQL Optimisation
  • Transaction Isolation & Locking

Cloud

  • AWS
  • EC2
  • S3
  • Bedrock

DevOps & Reliability

  • Docker
  • Kubernetes
  • KEDA
  • GitHub Actions
  • Custom Actions / Octokit
  • CI/CD
  • Linux
  • Chaos Mesh
  • k6

Observability

  • OpenTelemetry
  • Prometheus
  • Grafana
  • Jaeger / Grafana Tempo

Developer Tooling & Static Analysis

  • TypeScript Compiler API
  • ts-morph
  • Dependency Graph Analysis
  • Git History Mining

ML

  • TensorFlow
  • Scikit-learn
  • OpenCV
  • Model Evaluation

Education

2021 – 2025Vellore Institute of TechnologyB.Tech, Computer Science and Engineering · Vellore, India
Pre-UniversityDeeksha CollegePre-University College · Bengaluru, India

Let’s talk systems.

Questions about the work, ideas worth discussing, or something you’re building that has to stay up. Every message gets a reply.

email ajaynaidu5757@gmail.com · Bengaluru, Karnataka, India

Ajay Naidu · 2026Bengaluru, India