Abhishek Gupta
Software Engineer
Software Engineer
A collection of projects I've built over the years.
An advanced, AI-powered interview preparation platform with real-time emotion detection that helps candidates practice and improve their interview skills.
A privacy-first AI file organizer that classifies and structures files entirely on-device via a 3-layer pipeline — rule-based matching, TF-IDF + k-means clustering, and a quantized transformer (all-MiniLM-L6-v2) in WebAssembly. Zero uploads.
Real-time fraud-ring detection that models accounts, devices, and transactions as a Neo4j graph. Multi-hop Cypher traversal with union-find clustering achieved 100% recall and zero false positives on seeded rings.
An IR pipeline that ranks the top 100 best-fit candidates from 100,000 profiles in ~27s, CPU-only. Aspect-based retrieval + Reciprocal Rank Fusion + cross-encoder reranking reach NDCG@10 of 0.952 while catching 80+ honeypot profiles.
A production-ready RAG-based web application that allows users to upload books and have context-aware, citation-backed conversations with an AI assistant.
A DevOps framework that applies engineering discipline — testing, diffing, gating, and reverting — to AI agent memory built on Cognee knowledge graphs. Cut retracted-secret leakage from 12/12 adversarial probes to 0/12.
Lets AI agents measure their own reasoning by reading trace data — a LangGraph agent instrumented with OpenTelemetry, with SigNoz Trace Funnels exposing the plan→tool→validate→respond conversion rate through custom MCP tools.
A Slack agent that detects solo late-night firefighting in incident channels and offers consent-gated intervention: a supportive DM, an AI-drafted handoff note, and automatic paging of a rested teammate via a PagerDuty MCP server.
A dependency health agent that scans repos for NPM, PyPI, and Cargo manifests, matches installed versions against the OSV vulnerability database, and surfaces results through a risk-heatmap dashboard and a conversational Telegram bot.
A unified developer dashboard aggregating GitHub PRs, Gmail, Calendar, Todoist, weather, and dev articles, with a Gemini-powered assistant that answers natural-language questions by querying 7+ sources simultaneously through Coral's federated SQL engine.
A Python sidecar that turns Home Assistant automation traces into OpenTelemetry span trees and exports them to SigNoz as flame graphs — making automation behaviour legible where the built-in trace view falls short.
Diagnosing AI-agent performance with OpenTelemetry and SigNoz: an agent taking 21 seconds to answer "What is 2 + 2" is traced down to a silent-retry-loop bug and fixed to run in ~2.6 seconds.
A high-performance UCI-compliant chess engine in C++20 built for competitive play on Lichess — magic-bitboard move generation and a full modern search stack (alpha-beta/PVS, iterative deepening, transposition tables) with NNUE evaluation.
A Trello-style Kanban board (Laravel 12 + React 19) notable for how it was built: an autonomous two-agent AI system where a planning agent and a coding agent collaborate over Slack to design and implement the app.
An AI-augmented mentorship platform pairing mentees with expert mentors for live WebRTC video sessions, backed by an AI mentor that answers questions using retrieval over past successful mentorship conversations.
A customer credit-ledger PWA for village shopkeepers — offline-first tracking of who owes what, built with React, Vite, and Supabase, installable with Excel export support.
A product catalog and admin dashboard for an electrical-goods retailer — React + Vite + shadcn storefront backed by a separate Express/MongoDB API. Client work.
An AI recipe generator built on Next.js 15 and Google's Genkit that turns available ingredients into complete recipes using Gemini.
A browser-based photo framing and editing tool built with Next.js that lets users add frames and adjustments to their images entirely client-side.
A personal Discord bot that answers messages on the owner's behalf using Azure OpenAI (GPT-4-32k), forwards DMs to a channel, and pings the owner on name mentions.
