CS undergraduate focused on backend engineering and system design — building APIs, data pipelines, and distributed systems that are correct, observable, and built to scale.
I care about the full stack below the UI: schema design, async processing, algorithmic correctness, and infrastructure that doesn't require babysitting.
| 💻 Backend | 🗄️ Data & Databases | 🤖 ML Systems | ☁️ Infrastructure |
|---|---|---|---|
| REST APIs · FastAPI · Go | PostgreSQL · Redis · SQLAlchemy · MongoDB | XGBoost · Pipelines · Feature Eng. | Docker · GitHub Actions · AWS · Azure |
Football Market Intelligence — Sports Analytics & Prediction Engine
A quantitative sports analytics platform that identifies mispriced betting markets for the FIFA World Cup by comparing bookmaker-implied probabilities against proprietary ML-driven probability models.
What's under the hood
- Multi-schema PostgreSQL design (
raw→core→features→predictions→opportunities) with Alembic-managed migrations for deterministic deployments - FastAPI backend following domain-driven design: repositories handle DB queries, services hold business logic, API layer exposes typed endpoints
- XGBoost multi-class classifier (
multi:softprob) trained on Elo ratings, squad market value, rest days, and H2H history; Logistic Regression baseline for anomaly detection - Monte Carlo tournament simulation (10k–100k runs) with dynamic re-simulation on real-world result updates to reprice futures markets in near-real-time
- Background polling loops for API-Football (fixtures, lineups, injuries every 6h; odds every 5–15 min) with a robust identity resolution layer for cross-provider team name normalization
- Kelly Criterion bankroll allocation + EV-based filtering with liquidity, volatility, and max-exposure guards
- Async SQLAlchemy (
AsyncSessionLocal) + Uvicorn ASGI for high-concurrency request handling - Next.js + TypeScript frontend with live dashboards: Command Center, Match Analysis, Simulation Hub, and Performance Analytics
- Dockerized multi-service stack with Prometheus integration for system observability
ClassSync — Automated Timetable Generation System
A full-stack scheduling platform that automatically generates conflict-free timetables for schools and universities, solving a multi-dimensional constraint satisfaction problem under real institutional rules.
What's under the hood
- Custom scheduling engine combining heuristic prioritization, greedy placement, and Min-Conflicts local search — runs multiple randomized attempts and selects the solution with the highest placement score
lesson_weightheuristic ranks lessons by scheduling difficulty (day restrictions, faculty load, double periods, constraint count) before placement begins- O(1) conflict detection via multi-dimensional in-memory occupancy matrices (
class_occ,faculty_occ,room_occ); eviction-and-requeue loop resolves deadlocks without backtracking restarts - Supports rich constraint types: hard exclusions, max-one-per-day, subject sequences, same-day exclusions, day restrictions, and first-period reservations
- FastAPI backend with Pydantic schemas for validation; services layer cleanly separated from routes and DB repositories
- Async task offloading via Celery + Redis broker, keeping the API responsive during computation-heavy generation runs
- PostgreSQL with SQLAlchemy ORM; strict referential integrity enforced via foreign keys and cascade operations across all timetable entities
- Single Page Application frontend (TypeScript + Vite) with an admin dashboard to configure teachers, classes, subjects, and constraints
- One-command Docker Compose deployment (
docker compose up --build)
ConfessIt — Anonymous Social Platform
A full-stack social platform built for anonymous confessions, algorithmic matchmaking, and direct messaging — designed for an academic community with a heavy focus on backend architecture, observability, and security.
What's under the hood
- Layered FastAPI backend (routers → services → data access) with Pydantic v2 for strict schema validation at every boundary; auto-generated OpenAPI docs via Swagger
- Passwordless Magic Link authentication: secure tokens hashed and stored in MongoDB, verified against JWTs issued on success — no passwords ever stored
- MongoDB document store with startup-time index creation (
idx_timestamp_desc,idx_user_regno) for sub-millisecond querying on confessions, matches, and messages; Redis for session caching and transient matchmaking state - Feature surface: anonymous confessions with emoji reactions and threaded comments, algorithmic matchmaking, love notes with Cloudinary image storage, in-app conversation threads, and a full admin moderation dashboard
- Role-based access control (
user/admin) enforced via FastAPI dependency injection on every protected route;AdminRoutewrapper prevents unauthorized UI access - Full production observability stack: OpenTelemetry auto-instrumentation → collector → Grafana Tempo (traces), Promtail → Loki (logs), Prometheus + Grafana (metrics) — all provisioned via Docker Compose
- TanStack Query on the frontend for server-state caching and optimistic invalidation; React Hook Form + Zod for type-safe client-side validation
- Pytest +
mongomockfor backend unit testing; Vitest + React Testing Library for frontend component tests
Continumm — Network Telemetry & Observability Platform
A production-grade network monitoring backend that auto-discovers devices on subnets, polls health continuously, and exposes results through a REST API, Prometheus metrics, and pre-built Grafana dashboards.
What's under the hood
- Flask + Gunicorn backend with a dedicated background thread running an asyncio event loop for network I/O — discovery and health polling run independently without ever blocking the API
- Multi-source device discovery using ARP, Scapy, and Nmap ping sweeps across configured subnets; merges results by IP and upserts into PostgreSQL, with optional concurrent service fingerprinting (
nmap -sV) per discovered host - Continuous health polling every 30s via ICMP and HTTP probes; parses latency, jitter, and packet loss per device — writes time-series
DeviceStatusrows and triggersAlertEventrecords when configurable thresholds are breached - PostgreSQL advisory locks for distributed leader election — only one telemetry worker activates across horizontal replicas, guaranteeing single-scanner behavior without any external coordinator like Zookeeper or etcd
- REST API with endpoints for device inventory, historical metrics, alert events, and a telemetry overview; Flask middleware injects request IDs and emits structured JSON logs with trace IDs for cross-signal correlation
- Nginx reverse proxy for rate limiting, security headers, and routing; backend never exposed directly to the host with all config (DB credentials, scan targets, thresholds) driven by environment variables per 12-factor principles
- Full observability stack: OpenTelemetry auto-instrumentation → Tempo (traces), Promtail → Loki (logs), Prometheus → Grafana dashboards — entire 9-container stack provisioned via Docker Compose
- Terraform (AWS) provisions EC2 instances, VPC, and security groups;
cloud-init.yamlhandles Docker setup and firewall enforcement on first boot with zero manual steps - Kubernetes deployment via Kustomize with discrete Deployments, Services, and Nginx Ingress; multi-stage Docker build injects the Git commit for deployment traceability
- Pytest suite spanning API routes, config, discovery logic, models, and the telemetry service; live endpoint verification script confirms stack health post-deployment
