Utkarsh
Mittal.

Off-the-shelf RAG systems hallucinate without warning. Users get confident-sounding wrong answers with no way to verify them. This system traces every claim back to an exact source span and quantifies how sure it is.

In regulated industries like legal, finance, and healthcare, a wrong answer with a citation is worse than no answer. Grounded retrieval with measurable precision lets companies actually deploy AI on internal documents.

Hybrid BM25 and vector retrieval fused via RRF, cross-encoder reranking, SOURCE_X citation injection into the LLM prompt, then claim-level confidence scoring via cosine similarity against cited chunks.

High-throughput event streaming is critical infrastructure for modern data pipelines, microservice communication, and real-time analytics. Most teams use Kafka as a black box with no visibility into what happens at the storage and replication layer.

When Kafka goes wrong at scale, engineers who understand segment files, ISR shrinkage, and high-watermark semantics fix it in hours. Engineers who only know the client API spend days. This project demonstrates that depth.

Segment files on disk with sparse index files for O(1) seeks, a custom binary TCP wire protocol, ISR replication with high-watermark tracking, and a consumer group coordinator with JoinGroup and SyncGroup state machine.

Collaborative software requires multiple users to edit shared state simultaneously without conflicts, lost updates, or stale reads. Traditional locking breaks under network partitions and does not support offline editing.

CRDTs power the real-time layer in Notion, Figma, and Linear. Understanding the algorithm well enough to implement a WebSocket sync server and offline-first client is a meaningful signal for product-facing engineering roles.

Custom WebSocket server with Yjs binary sync protocol, IndexedDB offline persistence on the client, Postgres snapshots every 30 seconds for durability, and cursor presence via the Awareness protocol.

Every API needs rate limiting. Most implementations use a naive check-then-set pattern that has race conditions under concurrency: two simultaneous requests can both read under limit and both get through, breaking the contract.

Rate limiting is security-critical infrastructure in any multi-tenant SaaS product. Getting the algorithm and atomicity right under load, then proving it with benchmarks and IaC, is the kind of ownership companies look for in senior engineers.

Three algorithms as atomic Redis Lua scripts with zero race conditions, per-tenant config with 30-second in-memory cache and Redis pub/sub invalidation, Terraform-deployed to Fly.io, and load-tested to 8K+ RPS with k6.