I build AI-powered systems with production-grade backend architecture. Currently shipping RAG pipelines, hybrid vector search, and real-time market intelligence.
class Pavan:
role = "Full-Stack Developer & AI/ML Engineer"
education = "B.Tech CS (AI & ML) — VVIT, Graduating May 2026"
certifications = [
"Google Cloud Associate Cloud Engineer",
"AWS Certified Cloud Practitioner",
]
achievements = [
"1st Place — ACM Programming Contest (200+ participants)",
"Open Source Contributor — Wikimedia Foundation",
]
currently_building = "AI-powered GitHub Repository Analyzer"
fun_fact = "I optimized a RAG pipeline from 9.6s to 0.48s and it felt better than any game win"
A full-stack system that generates personalized, multi-phase learning paths using RAG, hybrid vector search, and U.S. labor market data. Not a wrapper around ChatGPT — a complete AI pipeline with measured performance.
🔍 Technical Deep Dive (click to expand)
L0 In-Memory (Python dict) → ~0ms — role + query tuple key
L1 pgvector Hybrid Search → 1 DB trip — HNSW cosine + B-tree metadata filter
L2 Tavily Live Web Fetch → 1-3s — only for cache misses, parallelized
Measured Performance:
| Scenario | RAG Latency | Total Generation |
|---|---|---|
| Cold start (0 cache hits) | 9.6s | 18.1s |
| Warm (L0 in-memory hit) | 0.48s | 10.1s |
| Improvement | 95% reduction | 44% reduction |
- 1024-dim Mistral embeddings with HNSW cosine index
- B-tree metadata filter on
target_rolepartitions search space per role - N vector lookups batched into 1 SQL query via
UNION ALL - Language-aware cache bypass: non-English queries always fetch fresh results
pg_try_advisory_xact_lock(user_id)prevents duplicate generation from concurrent requests- Per-user locking — 100 users generate paths fully in parallel
- Server-side test answers — MCQ correct answers never sent to frontend
- Fuzzy matches roles to SOC codes across 900+ occupations
- Extracts Hot Technology skills, knowledge domains, work activities
- LLM fallback for modern roles absent from O*NET
| Project | What it does | Stack |
|---|---|---|
| NextVentures | Startup discovery platform with SSR, GitHub OAuth, and real-time CMS sync | Next.js 14, TypeScript, Sanity, NextAuth |
| HandShake | Real-time ASL gesture recognition — CNN + MediaPipe hand landmark detection | TensorFlow, OpenCV, MediaPipe, Flask |
| PrepWise | AI interview prep with real-time speech analysis and feedback | Node.js, FastAPI, React, Socket.io |
Jul 2024 – Sep 2025 · Remote
- Built community insights dashboard (Python + Streamlit + MySQL) surfacing editor contribution trends
- Reduced SQL query execution time ~25% via composite indexing on replicated production datasets
- Authored reusable Python query modules with unit tests, eliminating ad-hoc analysis scripts
- Contributions merged after international code reviews under Wikimedia engineering standards
🥇 1st Place — ACM Programming Contest (200+ participants)
Building things that work at scale — not just things that compile.