All Projects

Keyword job boards flood candidates with noise while the roles that actually fit their experience get buried. Matching by keyword overlap fails to capture whether a job truly aligns with a résumé, and single-user automation scripts cannot serve multiple users with isolated, private data.

Treated matching as a semantic retrieval problem: embedded both résumés and job descriptions into the same 768-dimension Gemini vector space and ranked every job by pgvector cosine similarity. Built an idempotent n8n ingestion pipeline pulling from Adzuna and RemoteOK daily with SHA-256 URL-hash deduplication, and enforced true multi-tenancy with Postgres Row-Level Security and asymmetric ES256/JWKS JWT verification.

Shipped a live multi-user product self-hosted on a DigitalOcean droplet behind a Cloudflare Tunnel (zero open inbound ports) with the Next.js frontend on Vercel. First live ingestion pulled 72 real jobs into the shared pool; matching, auth isolation, dedup, ingestion, and the opt-in nightly email digest were each verified end to end.

Traditional Retrieval-Augmented Generation (RAG) pipelines suffer from a "Context Tax"—bloating LLM input with noisy documents which spikes latency and increases hallucination risks. Furthermore, strict, sequential one-way pipelines prevent adaptive backtracking when early retrieval steps fail.

Formalized the RAG process as a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) driven by a learnable Orchestrator policy. Implemented a Shared Global Working Memory (SGWM) to prevent context drift and established a "Confidence as Currency" bi-directional negotiation protocol.

Slashed the over-search rate by 91% (dropping from 27% to 2.3%), eliminating unnecessary retrieval rounds. Improved exact match and F1 scores significantly while achieving a 42% faster average inference time.

Existing technical hiring platforms face functional fragmentation, scalability constraints under high concurrent user loads, and static assessments that fail to adapt in real-time to candidate performance.

Co-architected a full-stack low-latency platform using the MERN stack and Google Gemini API. Engineered a multi-agent MockLLM framework to dynamically adjust questioning, backed by Redis caching and BullMQ job queues for async orchestration. Integrated WebRTC and Socket.io with Monaco Editor for a real-time collaborative workspace.

Delivered bidirectional code synchronization with <70ms latency and a multi-language code execution engine boasting a 1.4s average response time. Reduced overall API latency by 46% (890ms to 420ms) and improved candidate answer quality by 12.3% during pilot evaluations.

Managing data ingestion for multiple clients requires strict tenant-level isolation, scalable data partitioning, and robust error handling to prevent manual debugging bottlenecks.

Built a multi-tenant ETL system utilizing Apache Airflow and PostgreSQL with isolated, JSON-driven DAG configurations. Integrated the Groq API to automatically parse system logs and classify pipeline failures, enforcing a two-phase execution flow to validate AI suggestions.

Successfully processed over 100K records per day while cutting new-tenant onboarding time to under 5 minutes. Reduced mean pipeline resolution time from roughly 30 minutes to under 10 minutes with an 85%+ accuracy rate in failure classification.

Querying large, unstructured document corpora requires an architecture capable of handling concurrent ingestion and retrieval requests without blocking system throughput.

Designed a highly concurrent retrieval-augmented generation pipeline using FastAPI and Python. Implemented document chunking, embedding, and optimized index writes to a Pinecone Vector Database, packaging the entire environment into Docker containers with configurable worker counts.

Achieved consistent retrieval latency and relevance scores under high concurrent load in a production-like environment.