● Software Engineer (Full-Stack & Backend)
Sumit Parmar
I build complete systems end to end: Java and Spring Boot microservices with Kafka, React and TypeScript frontends, PostgreSQL, and cloud deployments on AWS, and I prove they work with tests, load tests and live demos.
MS Computer Science, University at Buffalo · December 2026 · Open to Software Engineer, Full-Stack and Backend roles
Live system map
- PharmaVigilchecking…AI drug-safety (pharmacovigilance) platform
- Nivesh-Pathchecking…AI behavioral-finance paper-trading platform
- Unix-style OS kernelCI + benchmarksPintos x86 kernel in C, extended into a sandbox
How each system moves data:
- PharmaVigil: Report → Kafka → Models → E2B · ACK
- Nivesh-Path: React → Express → Ledger → AI coach
- Unix-style OS kernel: syscall → umem → VM → swap
01 / Systems
Three systems, built end to end and measured.
Each one has a case study: the problem, the architecture you can click through, one request followed from start to finish, the decisions and their trade-offs, the hardest bug, and results linked to the files that prove them.
- SYSTEM 012026checking…
PharmaVigil
AI drug-safety (pharmacovigilance) platform
Turns a pharmacist's side-effect report into a reviewed, regulator-ready ICH E2B(R3) submission, and brings signals from real FDA data back to the counter.
Designed and built solo · Not for clinical or regulatory use
- Java 21
- Spring Boot
- Kafka
- JMS (Artemis)
- Oracle
- React
- TypeScript
- ONNX Runtime
- Kubernetes
- unique cases after de-duplication
- 10,848,477
- p95 report-to-coded after both fixes
- 7.5 s
- AUC, best classic signal method (IC) on OMOP
- 0.821
- SYSTEM 022025 – 2026checking…
Nivesh-Path
AI behavioral-finance paper-trading platform
Lets people practise investing with virtual cash, then detects their own biases (panic selling, FOMO buying, overconcentration) from their trades and coaches them.
Designed and built solo · Paper trading · educational, not financial advice
- React
- TypeScript
- Node.js
- Express
- PostgreSQL
- Prisma
- FastAPI
- LangChain
- AWS
- Terraform
- tests (Jest, Supertest, pytest, Vitest, Playwright)
- 300+
- faster read p95 with the cache
- 2.3×
- $10,000 buys against real PostgreSQL; exactly 10 fill
- 20 parallel
- SYSTEM 032025 – 2026
Unix-style OS kernel
Pintos x86 kernel in C, extended into a sandbox
Runs many untrusted programs safely on one machine: isolated memory, bounded resources, real pipes, and demand paging for working sets larger than RAM.
Course projects 1–2, then extended solo (sandbox, VM, fuzzing) · Coursework-based (Stanford Pintos)
- C
- x86
- QEMU
- Docker
- Python
- GitHub Actions
- CodeQL
- fuzzed system calls
- 116,000
- tests passing in CI across six suites
- 245
- fewer swap-ins with clock than FIFO
- about 100×
02 / Experience
Where I've worked and studied.
May 2026 – Aug 2026
Full Stack Development Intern · Uplifty AI
Texas, remote
- Developed cross-platform mobile features (user profiles, social connections and the challenge flow) for a consumer social app in React Native (Expo) and TypeScript, turning Figma designs into reusable UI components.
- Designed PostgreSQL schemas on Supabase with row-level security policies that enforce per-user access inside the database, and built serverless backend logic with Supabase Edge Functions in TypeScript.
- Integrated an LLM-powered feature and shipped each feature through Git feature branches, peer-reviewed pull requests and Jira tickets in a distributed Agile team.
Jan 2025 – May 2025
Software Engineer Intern · Prodigy InfoTech
Remote
- Designed and developed full-stack applications with React, Next.js, TypeScript, Node.js and PostgreSQL, implementing RESTful APIs, authentication and modular backend services.
- Built and maintained a regression suite of 450+ automated test cases (Jest, Supertest, Playwright), cutting regression execution time from 8 hours to 25 minutes.
- Built CI/CD pipelines with GitHub Actions, Docker and AWS, and optimized PostgreSQL queries with indexing, query analysis and schema improvements to cut response times.
Apr 2024 – May 2024
Full Stack Development Intern · Prodigy InfoTech
Remote
- Built responsive web applications in JavaScript, HTML5 and CSS3: a landing page with light/dark mode and sign-in, a stopwatch, a tic-tac-toe game and a weather app backed by a REST weather API.
- Refactored sequential API calls to async/await with parallel request batching, removing redundant round-trips on first render.
Jun 2023 – Jul 2023
Web Development Intern · Green Apex Solutions Limited
Ahmedabad
- Built reusable React.js components in a component-driven architecture and introduced a Redux store to replace prop drilling, giving related components a single source of truth.
Education
MS Computer Science (Systems Track)
University at Buffalo
Jan 2025 – Dec 2026 · GPA 3.5
B.Tech Computer Engineering
Indus University
Jun 2021 – Dec 2024 · GPA 9.08/10
03 / Skills → Proof
Not a list of logos: where each skill was used, and what it solved.
Filter by area. Every row links to the file that proves it.
Showing 43 uses.
| Skill | Used in | What it solved | Proof |
|---|---|---|---|
| Java Backend | PharmaVigil | 5 services on Java 21, with 106 tests and a coverage gate per service | test report ↗ (proof for Java in PharmaVigil) |
| Spring Boot Backend | PharmaVigil | intake, case processing, submission, pharmacy and signal services, each with its own schema and outbox | services/ ↗ (proof for Spring Boot in PharmaVigil) |
| Kafka Backend | PharmaVigil | reliable hand-off between pipeline stages with no lost reports: transactional outbox, idempotent consumers, 3 retries, dead-letter topics | ADR-001 + load test ↗ (proof for Kafka in PharmaVigil) |
| JMS Backend | PharmaVigil | the regulator hand-over: one E2B(R3) message out, one ACK or NACK back, recorded once under a row lock | ADR-002 ↗ (proof for JMS in PharmaVigil) |
| Node.js / Express Backend | Nivesh-Path | a REST API with 40+ endpoints, JWT on private routes, server-side pricing and idempotent orders | src/ ↗ (proof for Node.js / Express in Nivesh-Path) |
| FastAPI Backend | Nivesh-Path | the AI service: metrics computed in Python, retrieval, schema-validated answers, and a rule-based fallback | ai-service/ ↗ (proof for FastAPI in Nivesh-Path) |
| JPA / Hibernate Backend | PharmaVigil | the duplicate-check bottleneck: one query per candidate became one batched query (p95 103 s → 21 s) | PERFORMANCE.md ↗ (proof for JPA / Hibernate in PharmaVigil) |
| React Frontend | PharmaVigil | pharmacy counter, reviewer console and signals page for two roles | frontend/ ↗ (proof for React in PharmaVigil) |
| React | Nivesh-Path | trading, portfolio, Behavioral Mirror and AI advisor screens | frontend/ ↗ (proof for React in Nivesh-Path) |
| TypeScript Frontend | Nivesh-Path | typed API responses and UI state across the single-page app | frontend/ ↗ (proof for TypeScript in Nivesh-Path) |
| TypeScript | PharmaVigil | typed reviewer and pharmacy screens over five services' APIs | frontend/ ↗ (proof for TypeScript in PharmaVigil) |
| PostgreSQL Data | Nivesh-Path | a ledger that can't overdraw: NUMERIC money, CHECK constraints, SELECT … FOR UPDATE; 20 parallel orders, never overdrawn | concurrency test ↗ (proof for PostgreSQL in Nivesh-Path) |
| Oracle Data | PharmaVigil | the primary database, with the same test suite also green on SQL Server | ADR-003 ↗ (proof for Oracle in PharmaVigil) |
| MongoDB Data | Nivesh-Path | users, Behavioral Mirror events and the source of truth for rebuilding AI memory | src/models/ ↗ (proof for MongoDB in Nivesh-Path) |
| Redis Data | Nivesh-Path | read p95 22.3 ms → 9.6 ms and a 99.7% hit rate, keeping inside the market-data rate limit | load-test results ↗ (proof for Redis in Nivesh-Path) |
| PySpark Data | PharmaVigil | 12,902,877 FAERS rows into 10,848,477 unique cases through a bronze → silver → gold pipeline | data_counts.json ↗ (proof for PySpark in PharmaVigil) |
| Python Data | PharmaVigil | signal statistics (PRR, ROR, IC, EBGM) with bootstrap confidence intervals on two reference sets | results ↗ (proof for Python in PharmaVigil) |
| Python | Kernel | the benchmark analysis and a trace-driven simulator comparing eviction policies | SIMULATION.md ↗ (proof for Python in Kernel) |
| ONNX Runtime AI | PharmaVigil | three models served inside Java with no extra service, kept identical to Python within 1e-5 | ADR-008 ↗ (proof for ONNX Runtime in PharmaVigil) |
| LangChain / RAG AI | Nivesh-Path | coaching that quotes the user's own trades, from a per-user vector index rebuilt from MongoDB after every deploy | architecture.md ↗ (proof for LangChain / RAG in Nivesh-Path) |
| AWS Cloud/DevOps | Nivesh-Path | ECS Fargate (1–3 tasks), RDS PostgreSQL, CloudWatch, and an EventBridge → Step Functions → Lambda nightly job; created and destroyed to save cost | AWS.md ↗ (proof for AWS in Nivesh-Path) |
| Terraform Cloud/DevOps | Nivesh-Path | the AWS stacks, deployed from GitHub with OIDC and no stored keys; Checkov in CI | infra/ ↗ (proof for Terraform in Nivesh-Path) |
| Terraform | PharmaVigil | AKS, ACR and Key Vault for Azure: validated and Checkov-clean, never applied | AZURE.md ↗ (proof for Terraform in PharmaVigil) |
| Kubernetes Cloud/DevOps | PharmaVigil | the whole stack on kind with Helm and Strimzi Kafka, deployed and smoke-tested in CI | KUBERNETES.md ↗ (proof for Kubernetes in PharmaVigil) |
| Jenkins Cloud/DevOps | PharmaVigil | a pipeline as code (JCasC): tests, coverage gate, images, Helm lint, kind deploy, smoke test | JENKINS.md ↗ (proof for Jenkins in PharmaVigil) |
| Docker Cloud/DevOps | Kernel | a reproducible QEMU toolchain, so every suite runs the same in CI and locally | Dockerfile ↗ (proof for Docker in Kernel) |
| Docker | PharmaVigil | the demo stack fitting a free 12 GB server: 6,055 MiB peak | memory measurement ↗ (proof for Docker in PharmaVigil) |
| GitHub Actions Cloud/DevOps | PharmaVigil | 17 checks on every push, including an arm64 run of the demo stack | workflows ↗ (proof for GitHub Actions in PharmaVigil) |
| GitHub Actions | Kernel | every suite, the fuzzer, cppcheck and CodeQL on each push, gated on a regression baseline | workflows ↗ (proof for GitHub Actions in Kernel) |
| OpenTelemetry Cloud/DevOps | Nivesh-Path | one trace across the Node and Python services on the live site (26 spans) | live evidence ↗ (proof for OpenTelemetry in Nivesh-Path) |
| OpenTelemetry | PharmaVigil | trace context carried through the outbox: 182 spans in one trace | OBSERVABILITY.md ↗ (proof for OpenTelemetry in PharmaVigil) |
| Prometheus & Grafana Cloud/DevOps | PharmaVigil | the metrics that exposed both pipeline bottlenecks; three dashboards with alert rules | OBSERVABILITY.md ↗ (proof for Prometheus & Grafana in PharmaVigil) |
| k6 Testing | PharmaVigil | found two bottlenecks: p95 report-to-coded 103 s → 7.5 s, 0 failed | PERFORMANCE.md ↗ (proof for k6 in PharmaVigil) |
| k6 | Nivesh-Path | measured what the cache saves: 2.3× faster read p95 over 42,000 requests | PERFORMANCE.md ↗ (proof for k6 in Nivesh-Path) |
| Testcontainers Testing | PharmaVigil | integration tests against real Oracle, SQL Server and Kafka instead of in-memory stand-ins | e2e-tests/ ↗ (proof for Testcontainers in PharmaVigil) |
| Jest / Supertest Testing | Nivesh-Path | boundary tests at every Behavioral Mirror threshold, each fix broken on purpose to prove its test fails | report ↗ (proof for Jest / Supertest in Nivesh-Path) |
| Playwright Testing | Nivesh-Path | a live smoke test that signs up, trades and deletes a real account | live check ↗ (proof for Playwright in Nivesh-Path) |
| Playwright | PharmaVigil | a daily smoke test of the live demo: replay, ACK, E2B file, signals | workflow ↗ (proof for Playwright in PharmaVigil) |
| Fuzzing Testing | Kernel | 116,000 adversarial system calls, 0 panics; found a kernel panic on exec("") | fuzz record ↗ (proof for Fuzzing in Kernel) |
| C Systems | Kernel | system calls, virtual memory, pipes and resource limits in a Pintos x86 kernel | src/ ↗ (proof for C in Kernel) |
| x86 / QEMU Systems | Kernel | page tables, page faults and the int $0x30 system-call trap, tested under QEMU | SYSCALLS.md ↗ (proof for x86 / QEMU in Kernel) |
| Concurrency Systems | Kernel | two swap races and a use-after-free fixed with a condition-variable barrier and a lock-order rule | BUGS-FOUND.md ↗ (proof for Concurrency in Kernel) |
| Concurrency | Nivesh-Path | parallel orders from two app instances serialised by a row lock | test ↗ (proof for Concurrency in Nivesh-Path) |
04 / Other work
Course projects.
- Course project · Data Intensive Computing · 2026Cloud Financial Analytics PipelineA Databricks medallion ETL pipeline with MLflow experiment tracking and a PyTorch LSTM autoencoder for anomaly detection.DatabricksPySparkDelta LakeMLflowPyTorch
- Team of 3 · Deep Learning course · 2026In progressTransformer with PPO fine-tuningBuilding a small language model from scratch and fine-tuning it with proximal policy optimization (PPO).PythonPyTorchTransformersPPO
06 / Contact
Let's talk.
I'm graduating in December 2026 and looking for Software Engineer, Full-Stack and Backend roles. The fastest way to reach me is email.