Bid Engine
Building the intelligence behind construction tender estimation: document extraction, agent tooling, retrieval, and cloud-native event pipelines.
I'm Domenico Nota, a University of Toronto Computer Science student building backend systems, distributed infrastructure, and AI applications.
I'm a Computer Science student at the University of Toronto, with a minor in Mathematics. What drives me isn't just making software work. It's understanding complex problems deeply enough to make them feel simple.
From low-level infrastructure to AI-powered workflows, I like taking ideas from rough architecture to working systems that people can actually rely on.
The best architecture proves itself in production. Here's where my work meets the real world.
Building the intelligence behind construction tender estimation: document extraction, agent tooling, retrieval, and cloud-native event pipelines.
Owned the full stack of a real-time market simulation platform, connecting live-data ingestion, forecasting workflows, and a React experience.
Built an accessible, responsive React application for 700+ monthly users and helped ship it with automated deployment to AWS EC2.
Explore the architecture, implementation choices, and testing behind each project.
A local log pipeline with bounded Rust collection, Kafka buffering, idempotent PostgreSQL writes, and a React explorer.
Event-driven order, payment, inventory, and shipping services engineered to recover gracefully when things break.
Full-text search with global TF-IDF ranking, consistent-hash routing, persisted shards, and explicit partial results.
Collect application logs, absorb database interruptions, and make stored events searchable without silently advancing past failed writes.
A Rust agent sends bounded HTTP batches to a Go API. Kafka buffers accepted events; a separate Go worker owns persistence and offset commits. PostgreSQL deduplicates repeated event IDs.
At-least-once delivery begins after Kafka acceptance. The agent uses memory rather than a disk spool, so events can be lost before that boundary. Cursor pagination keeps query traversal stable.
The repository includes a deterministic end-to-end demo that checks one event through the pipeline and asserts that repeated delivery stores one row. No throughput claim is made here without a recorded benchmark run.
Select a stage to inspect its responsibility. This is an architecture walkthrough, not a live backend.
Tails log files, assigns stable event IDs, and sends bounded batches through HTTP. Retries reuse the same IDs; the in-memory queue is not a durable disk spool.
Read path: React explorer → Go query API → PostgreSQL. Poison records: Go worker → Kafka dead-letter topic.
Role: Code contributor · Repository: tusharra0/Fault-Tolerant-Order-System
Coordinate order, payment, inventory, and shipping when messages can be delivered more than once and individual stages can fail.
Four service domains own separate databases. Transactional outboxes publish events through RabbitMQ; idempotent consumers and Saga-style compensation handle repeated deliveries and downstream failures.
At-least-once delivery and eventual consistency avoid requiring one transaction across all services, but require deduplication, outbox workers, and explicit compensation logic.
The repository reports a 2026-08-08 Locust run on a GitHub-hosted Ubuntu runner: 20 users, two minutes, 2,370 requests, 19.905 requests/sec, 160 ms p95, and zero request failures. This measures API acceptance, not end-to-end order completion.
Local Docker Compose implementation with simulated domain services; not a claim of a live production deployment.
View order-system repository & measured results ↗Role: Code contributor · Repository: tusharra0/Distributed-Search-Engine
Search independently persisted document shards with consistent relevance scores and clear behavior when a shard is unavailable.
A Spring Boot coordinator routes documents using consistent hashing. Two parallel fan-out phases gather corpus statistics and score matches using global TF-IDF before merging the top results.
Two network phases improve scoring consistency but add a round trip. Each document has one owner: unavailable shards produce explicitly marked partial results, not replicated failover.
The repository provides healthy-cluster and shard-failure tests, plus a reproducible JMeter workload. Throughput and latency depend on the machine, dataset, and concurrency; no measured result is asserted here.
Languages and frameworks are means to an end. I choose the ones that make a system more useful, reliable, and clear.
Python · Java · SQL · FastAPI · Spring Boot · REST APIs
Interested in discussing a software engineering opportunity? Get in touch.
START A CONVERSATION