Harsh Upadhyay

AI-Powered Codebase Intelligence SaaS

A multi-tenant SaaS that securely indexes GitHub repositories and answers natural-language questions about unfamiliar codebases — grounded in exact files, symbols, and commits, not hallucinated.

• FastAPI backend with async SQLAlchemy 2.x + PostgreSQL, and a Redis Streams-based durable job queue with retry and recovery.
• Tree-sitter static analysis extracts symbols and a conservative call graph without executing repository code.
• A bounded AI agent (3 read-only tools, max 4 calls) validates every citation against server-collected evidence to prevent hallucination.
• 332 backend tests at 91% branch coverage; strict mypy & Ruff clean; Dockerized CI with vulnerability scanning.

• Python
• FastAPI
• PostgreSQL
• Redis
• Tree-sitter
• Qdrant
• React
• TypeScript

Live Voice-Based AI Interview Coach

A production-deployed, full-stack app that runs real-time voice mock interviews — transcribing, evaluating, and scoring answers turn by turn as you speak.

• Real-time WebSocket state machine: speech-to-text → adaptive LLM question generation → 4-dimension answer evaluation → text-to-speech.
• Resume-aware personalization parses uploaded PDFs/DOCX into a candidate profile that drives question generation.
• Fine-tuned a Qwen2.5-1.5B evaluator (LoRA via Apple MLX) through knowledge distillation, cutting average scoring error from ~5.3–7.4 points to 0.6–1.9 on a 0–10 scale.
• Diagnosed and fixed real production incidents — an OOM crash from an eagerly-loaded local model, a WebSocket idle-timeout behind a reverse proxy, and cross-site cookie issues.

• Python
• FastAPI
• WebSockets
• PostgreSQL
• Groq
• React
• Tailwind
• MLX

Schema-Agnostic Data Governance Platform

A full-stack data quality tool that ingests any tabular dataset — no fixed schema — runs it through a configurable rule engine, scores it transparently, tracks lineage, and surfaces it on a dashboard with SQL-backed reporting and CSV export for BI tools.

• 7 independently-testable, schema-agnostic quality rules — completeness, duplicate detection, data-driven validity, outlier detection, near-duplicate consistency, and schema-drift — rolled into a transparent, itemized 0–100 health score.
• Hand-written parameterized SQL — window functions and CTEs for per-dataset run summaries, aggregate queries for issue breakdowns — benchmarked via EXPLAIN ANALYZE at 2.2ms and 16.3ms against 2,000 synthetic runs.
• Verified schema-agnostic by running the same unmodified code against two structurally unrelated datasets with zero config changes; OCR ingestion (pytesseract + pdf2image) for scanned tabular reports.
• Built a Power BI report on the pipeline's health-score trend and a validated Airflow DAG (TaskFlow API) alongside the live APScheduler automation; 67-test CI on every push via GitHub Actions.

• Python
• FastAPI
• PostgreSQL
• SQL
• Alembic
• Airflow
• Power BI
• React
• Docker

Enterprise Full-Stack Application

A deployed visitor management system with real-time notifications and automated reporting for multi-location organizations.

• 10+ REST endpoints for visitor check-in/out, approvals, and admin workflows on Node.js/Express with MongoDB.
• Push notifications to admins on new check-ins via Firebase Cloud Messaging.
• Automated daily Excel exports via a scheduled cron job, plus photo capture/upload via Cloudinary.
• Diagnosed and resolved data-integrity issues across concurrent registrations and approvals.

• React
• Node.js
• Express
• MongoDB
• JWT
• RBAC
• Docker