Abhinav Aditya

Commercial work at HireQuotient, a B2B recruiting platform. I built the outbound email engine from the first commit and still own it. It dispatches over 400,000 emails a day and handled more than 1.5 million in its busiest week.

This is the piece of work I point at when someone asks whether I can build something that survives real load.

The problem

Sending a lot of email is easy. Sending a lot of email that reaches inboxes is not. Providers rate limit per mailbox, reputation is per sender, and a naive queue that fans out as fast as it can will get an entire domain blocked inside a day.

How it works

Queue architecture. A horizontally scalable RabbitMQ and Redis system. Workers pull from provider specific queues, so a slow or degraded provider backs up its own queue instead of the whole system. Retries and dead lettering are built in rather than bolted on.

Sender pool rotation. Sends are distributed across a rotating pool of connected mailboxes with per mailbox pacing, which is what keeps deliverability intact at this volume.

Multi provider. Gmail, Outlook, SendGrid and raw SMTP or IMAP, each behind OAuth 2.0 connection flows, all behind one interface so adding a provider does not mean touching the send path.

Observability. Grafana, Prometheus and Sentry, because at this volume you find out about a problem from a dashboard, not from a customer.

Related work on the same platform

• A multi agent GenAI pipeline on async FastAPI: three cooperating stages named Gatekeeper, Generator and Critic that score relevance and personalise outreach, replacing what had been manual lead sourcing.
• A conversational recruiter copilot on LLM tool calling, exposing more than six natural language actions.
• More than six ATS and data provider integrations behind a single abstraction layer, each with OAuth 2.0 and two way sync.
• Infrastructure: Docker across five services, ECS and ECR provisioning with Terraform and Ansible, and a branch locked CI/CD pipeline in GitHub Actions that blocks unauthorised production deploys across the org.

Stack

Python, FastAPI, Node.js, Celery, RabbitMQ, Redis, AWS SQS, MongoDB, Docker, AWS ECS and ECR, Terraform, Ansible, GitHub Actions, OAuth 2.0, Grafana, Prometheus, Sentry

A resume tailoring SaaS that will not invent your experience. Most AI resume tools happily fabricate projects. Resumate solves that by grounding generation in evidence the candidate actually owns.

Live: https://www.twine.net/signin
Source: https://www.twine.net/signin

What it does

GitHub grounded knowledge base. A candidate connects their GitHub account and picks repositories. LangChain turns those repositories into a per user knowledge base, and every tailored resume is generated against that evidence rather than model invention.

Multi stage, multi provider pipeline. The first version was a single call to Gemini, which produced fluent and unreliable output. I replaced it with a multi stage pipeline across more than one provider, so drafting, grounding and scoring are separate steps that can each use the model best suited to them.

Product surface. An npm workspaces monorepo in React, Vite and Tailwind, with a multi step tailoring workspace and ATS result visualisation.

Backend. Node.js and Express on SQLite, secured with JWT and Helmet, credits based monetisation, and zero downtime deploys through PM2 behind Nginx.

Stack

LangChain, React, Vite, Tailwind CSS, TypeScript, Node.js, Express, SQLite, GitHub API, Google Gemini, JWT, PM2, Nginx

Context as a service. Synapse indexes an entire codebase, including uncommitted work in progress, and serves it to AI coding tools such as Cursor, Claude and Copilot over the Model Context Protocol.

The CLI is published on PyPI as synapse-context-cli: https://www.twine.net/signin

How it is built

Seven services. A FastAPI API, a Rust and Tree-sitter parser, a Temporal worker, a Typer CLI published to PyPI, and a Next.js admin dashboard.

Three tier retrieval. A Redis hot layer holding work in progress, merged with Qdrant branch and global namespaces under an override precedence rule. An uncommitted change wins over what is on the branch, and the branch wins over the global index. This is the part that makes answers match the code you are actually looking at rather than the code that was last merged.

Ingestion. GitHub push webhooks trigger indexing, Temporal orchestrates the workflow, and OpenAI embeddings back the vector search.

Security. Gitleaks scans every file before it can reach the embedding step and rejects anything carrying a secret, so credentials never land in the index. Per query ACL filtering on top, plus just in time org membership validation on a sixty second cache.

Supply chain. The CLI ships through GitHub Actions trusted publishing with Sigstore attestations.

Admin dashboard. Secret alerts, indexing status, and a live query playground.

Stack

Python, FastAPI, Rust, Tree-sitter, Temporal, Qdrant, Redis, PostgreSQL, OpenAI embeddings, MCP, Next.js, Docker, GitHub Actions, JWT

Built across 179 commits between February and July 2026.