def portfolio(): Abhay Munj

Orahee is a production travel-planning platform: a Django/DRF backend serving a Flutter app with AI-generated day-by-day itineraries, group collaboration, and real-time chat over Django Channels WebSockets — live with around 200 active users.

The itinerary engine keeps activities as relational rows while mirroring them into a flattened JSON document on the trip, so AI-generated and hand-built itineraries share one CRUD model. Trip generation runs through a Pydantic-validated pipeline enriched with Google Places geodata, and in-chat group polls feed their results straight into structured generation input. Firebase push notifications, Branch.io deep-link invites, and Celery-scheduled engagement flows round out the product.

An OCR pipeline that turns national trademark gazette PDFs into structured, validated JSON/CSV records. PyMuPDF extracts text and image bounding boxes from scanned gazettes, and layout-correction algorithms rebuild a coherent reading order from multi-column and tabular scans that would otherwise interleave unrelated records.

Every country onboards through a config-driven parser — regex section and record splitters plus a small per-country extractor — so a new registry needs a config file, not core changes. A per-record classifier routes ambiguous entries (renewal vs. cancellation) when gazettes print mixed action types with no section breaks, and a ~500 DPI cropping stage clips trademark logos out of the source PDFs, multi-threaded for throughput.

A retrieval-augmented generation API built to be measured, not just demoed. Retrieval is hybrid — FAISS dense embeddings fused with BM25 sparse search via Reciprocal Rank Fusion, then reranked with a CrossEncoder — and a reranker score threshold makes the system refuse to answer rather than hallucinate. Every answer carries page-level source citations.

An evaluation harness tracks hit-rate and MRR per pipeline stage plus LLM-judge faithfulness, with config snapshots for regression detection. A multi-turn tool-calling agent adds nested LangFuse tracing, and the whole stack ships as Docker Compose.

Production scrapers for e-commerce review mining and property-data aggregation across multiple high-volume sources, sustaining a ~95% success rate against dynamic JS-rendered pages, pagination, session management, and rate limiting.

Reliability comes from layered retry logic and Bright Data proxy rotation; correctness from normalizing everything into PostgreSQL schemas with automated validation that catches layout drift and silent extraction failures before they poison downstream data.

An AI cold-outreach platform that writes, sends, and manages sales conversations end to end. Each conversation is modeled as a LangGraph state machine with checkpoints persisted in PostgreSQL, so an in-progress thread pauses on a reply and resumes days later exactly where it left off.

Celery Beat dispatches new leads and fans out reply-checking, with per-lead Redis locks preventing double-processing. An LLM classifies inbound replies — interested, out-of-office, negative — and branches the workflow: book a meeting, schedule a human-like delayed follow-up, or halt outreach. User-supplied third-party credentials live in a Fernet-encrypted vault that exposes only a last-four-characters hint.