Shubham Tanwar
Backend & systems engineer.
I design and build systems end to end in Go, Rust, and TypeScript: event-driven platforms, real-time collaboration, and agent infrastructure. I’m 20 and focused on making people and AI agents work together in real time.
My work includes a multi-service SaaS platform with hybrid RAG and LLM classifiers, a calibrated cloud-architecture simulator, and an open-source multiplayer AI runtime. I document each system’s trade-offs and limits.
Download résumé (PDF)
Work
Founder · Apr 2026 – Oct 2026
A pre-launch platform for teams working with AI agents, built with Go, Rust, TypeScript, AWS, and Cloudflare.
- A Go monorepo of approximately 12 event-driven services builds into one static binary, with configurable deployment together or as separately scaled processes and roughly 750 test files across Go and TypeScript.
- Per-service SQS queues, dead-letter queues, and a PostgreSQL outbox support versioned events that are idempotent by event ID. Production refuses to boot when required queues are missing.
- Real-time collaboration uses Cloudflare Workers, Durable Objects, and WebSockets, with a guest-to-account handoff that preserves workspace context and HMAC-signed calls that purge private conversations when membership changes.
- Delegation compresses a team conversation into a typed brief, attaches authority when a human commits it, and runs per-user agents through durable Workflows and follow-up alarms.
- Hybrid RAG combines pgvector and PostgreSQL full-text search with reciprocal rank fusion and reranking. Corpora pin their embedding model and chunk policy; indexing uses the member’s own credentials.
- GitHub, Notion, Slack, Drive, and Gmail ingestion runs through Composio without service-held tokens. A Gemini judge filters index noise; a provenance gate and an evaluated privacy classifier check agent-to-agent answers.
- Concurrent agent output streams with per-workspace token budgets and Valkey-backed rate limiting. A Rust API prototype verifies EdDSA tokens against cached JWKS and reads roles at use time.
Open-source multiplayer AI runtime · Rust · Apache 2.0 · v0.7 pilot
Several people and coding agents—Claude Code, Codex, and OpenCode—work in one live room without modifying the CLIs. One-command installation includes browser and terminal views. Source code.
- An append-only event log supports late joining, replay, recovery, and audit; rooms survive restarts.
- Each agent streams on its own lane, a cursor into the event log.
- Declarative YAML connectors support ACP agents and ADK, CrewAI, LangGraph, and PydanticAI. A probe command runs a real turn before an agent joins a room.
Tectonic
Cloud architecture simulator · Rust, Go, and TypeScript
A deterministic discrete-event engine that simulates AWS components and Python, Node, and Go runtimes on one clock. Private repository; demo and code walkthrough on request.
- Models load balancers, RDS/Aurora, DynamoDB, Valkey, Lambda, SQS/SNS, autoscaling, and failover.
- A five-method ensemble reports where an architecture breaks and why as a range, streamed live over SSE.
- A Go importer turns OpenTofu/Terraform files or plan JSON into runnable models and exposes unknown traffic, CPU, and query-cost inputs as gaps.
- Calibrated against Neoverse N1 hardware: held-out levels with p99 predictions within ±25% improved from 2/19 to 13/19, with misses published alongside hits.
ComStack
Co-founder · Oct 2025 – Feb 2026
Multiplayer agent teams for commerce, taken to a paid pilot with 12 design partners.
- Built Synq: Go operations, commerce, and agent services on Cloud Run, with tenant-scoped PostgreSQL, pgvector for semantic catalog search and SKU classification, and Valkey for caching and rate limits.
- Temporal child workflows propagate tenant context in workflow headers; GCP Pub/Sub and Eventarc handle asynchronous events with dead-letter handling.
- Reduced inference cost with noise-reducing classification and open-weight models, then wound the pilot down when multiplayer usage outpaced what pricing could cover.
Xtride
Systems Architect (Contract) · Sep 2025 – Oct 2025
Built backend and spatial data systems for a Bangalore mobility startup handling real-time driver telemetry.
- H3/S2 spatial indexing for dispatch and ETA pipelines.
- Event-driven GPS processing and an idempotent trip state machine.
- Database pooling, query optimization, and edge caching for peak demand.
Engineering
I work across real-time collaboration, agent infrastructure, backend systems, and data-intensive products.
- Real-time systems: stateful multiplayer with Durable Objects and WebSockets/SSE; append-only event logs for replay, recovery, audit, and late joiners.
- Agent infrastructure: parallel streaming lanes, ACP and framework connectors, token budgets, Valkey-backed rate limiting, Temporal orchestration, and durable Workflows.
- Data and AI systems: PostgreSQL and PostGIS, hybrid RAG with pgvector, full-text search, RRF and reranking, credential-scoped indexing, and LLM classifiers with evals.
- Backend workflows: versioned events, transactional outboxes, dead-letter queues, idempotent trip state transitions, and tenant-scoped commerce operations.
- Simulation: deterministic discrete-event models, OpenTofu/Terraform imports, and calibration against measured hardware.
Current focus
I’m currently building the Slack integration for a Switchboard experience that connects team conversations with a shared workspace for people and AI agents.
Applied distributed databases at scale: sharding, storage engines, connection pooling, replication lag, and query optimization.
Languages: Go, Python, TypeScript, and Rust.
Data and infrastructure: PostgreSQL and PostGIS, Valkey/Redis, Cloudflare Workers, Durable Objects and Workflows, WebSockets/SSE, AWS (SQS, SNS, EventBridge), GCP (Cloud Run, Pub/Sub), OpenTofu/Terraform, vector search, and H3/S2.
AI systems: hybrid RAG, LLM classifiers and evals, Bedrock, Vertex AI/Gemini, open-weight models, LangGraph, ADK, ACP, and MCP.
About
I’m building Oryxa, a DPIIT-recognized company in India. My work spans real-time collaboration, agent infrastructure, and backend reliability.