Amit Singh.

Senior Backend Engineer · Applied AI · Pune, India

Backend systems.
Built for the failure cases.

I’m Amit Singh. I build and operate Python services on AWS, with a focus on data correctness, recovery from failure, and LLM components that stay under human review.

7+ years across financial data, network operations, and product-safety systems. Currently at Realtek Software Services, delivering backend, data, and platform features — including Bedrock-backed classification with confidence-gated review — on an Amazon Ring / Lab126 client engagement.

Download résumé PDFExplore selected workConnect on LinkedIn
Python & AWSTransactional correctnessProduction reliabilityApplied AI with human review

01 / Production experience

Ownership beyond the implementation.

Product-safety systems

Software Development Engineer II
Realtek Software Services · Amazon Ring / Lab126 client engagement
2025–present

Backend and platform delivery for a multi-company application, including data consistency, access boundaries, and operational recovery. I also build the React interfaces operators use.

  • Correct the owning service. Traced a cross-service authorization regression to an identity-resolution change and drove the correction at its source.
  • Recover from missed events. Designed compliance-sensitive reconciliation across UI updates, asynchronous events, and scheduled repair.
  • Make changes safer. Strengthened database-change controls with reviewed SQL, rollback plans, staged rehearsal, and post-change verification. Mentor two engineers and participate in on-call.

Client work is described at a high level. Source code and internal operational evidence are not public.

02 / Selected open-source work

Decisions you can inspect.

Independent projects, separate from client work. Each has a defined scope and links to its evidence.

Ledger & event-driven systems

CloudScale Backend

Keep ledger state correct when commands are retried and consumers restart.

Used transaction-bound idempotency, balanced postings, and a transactional outbox, with PostgreSQL and SQLite adapters behind shared contracts.

Evidence
PostgreSQL integration tests and an executable backup/restore drill that rebuilds derived state.
Boundary
Recovery measured at seeded scale; sustained availability is a target, not a demonstrated result.
Source & designTests & recovery evidence

Reliable delivery

FastAPI Microservices Platform

Self-hosted webhook delivery using PostgreSQL alone, without Redis or Kafka.

Leased jobs and token-checked updates prevent stale workers from overwriting a newer claim. Standard Webhooks signatures and isolated egress address the delivery boundary.

Evidence
A reproducible million-job correctness run recorded no lost jobs. The SDK is published separately on PyPI.
Boundary
At-least-once delivery; the backlog-latency target was not met. This is not a production-capacity claim.
Source & designBenchmark scope

Developer tooling

Code Quality Analyzer

Review changed code without sending source to an external service.

Offline analysis with SARIF reports, changed-line gates, and a GitHub Action. Configuration fingerprint checks can detect changes that weaken the agreed gate.

Evidence
Published as cqa-analyzer, with cross-platform CI and fixtures for reported findings.
Boundary
Bounded static analysis, not a substitute for data-flow security analysis or human review.
Source & rulesGitHub ActionValidation evidence
More work: agent workflows, transfers, and coding tools

AgentOS

Durable agent workflows over an event log, with approval gates and an operator UI. Crash tests exercise recovery; external effects still depend on downstream idempotency.

Public source & evidence

Agent Skills

Four tool-backed coding-agent workflows, including an architecture snapshot and a changed-line quality gate. Script tests are distinct from model-behaviour validation.

Scope & validation

03 / Technical leadership

Direction beyond the immediate fix.

Recurring failures become clearer boundaries, approved plans, and mechanisms that outlast the fix.

Architecture boundary

Retired cross-tenant coupling

Isolated Ring-specific criticality and ticket updates after shared Lab126 behavior caused defects, then made Project and Project Family capabilities company-aware end to end — closing the gaps that let one company’s values reach another’s workflows.

Roadmap & alignment

Turned opportunities into approved plans

Authored phased plans for source migration and automated criticality workflows, aligned stakeholders on the trade-offs, and secured approval to take source migration into detailed design.

Operational mechanism

Made database changes safer

After a production data-integrity incident: reviewed SQL, explicit rollback plans, staged rehearsal, and post-change verification for every later change and backfill. Risk is now caught before production instead of recovered afterward. I bring the same stance to open-source work and state where the evidence stops.

04 / Background

Backend first. Comfortable across boundaries.

Before Realtek

Orange Business Services · 2024–2025
Production CI/CD and a retrieval-grounded operations assistant, including user testing and operational handover.

Virtusa Polaris · 2019–2024
Financial-data APIs, PostgreSQL migration, and ETL for the Lloyds Banking Group engagement.

Working stack

Primary: Python, AWS, FastAPI, Django, SQL, PostgreSQL, event-driven services, CI/CD, and incident response.

Applied AI: Amazon Bedrock and Claude in production behind confidence-gated review, retrieval-grounded assistants with retrieval evaluation, and durable agent workflows with approval gates.

Supporting: React/TypeScript operator interfaces and hands-on Go projects.

Read the full résumé