github.com/kishorg1608 · newest first
Fifteen repositories, four years of shipping — from multi-agent platforms solving my own day job to the college projects where it all started. Every card links to the code.
An AI screening platform that parses resumes (PDF, Word, text) and evaluates them against a job description like a strict recruiter — semantic evaluation, structured JSON fit scores with reasoning, and candidate rankings. Keyword-stuffing tricks that fool ATS systems don't fool this.
The platform version of my day job. A LangGraph multi-agent system that automates Business Analysis and presales: discovery plans, minutes of meetings from transcripts, scope-gap analysis with SOW citations, and export-ready Agile user stories in Given/When/Then. Quality is enforced by actor–critic loops — a generator agent drafts, a reviewer agent judges, and work loops back on "REWORK" until it passes.
An intelligent translation companion for travelers and immigrants. Uses browser hardware APIs — camera, microphone, geolocation — for instant voice translation, sign reading through the camera, and location-aware context, with Gemini doing the heavy lifting behind a lightweight proxy server.
A real client site for a diagnostic centre — fast, responsive, and SEO-engineered for regional discoverability: meta targeting for local searches, XML sitemaps, robots schemas, and Open Graph previews. Vanilla HTML/CSS/JS on purpose, so it loads instantly for patients on any device.
A fully offline RAG platform — before it was cool. Upload PDFs and converse with them through a locally-run quantized Llama-2, so sensitive legal/medical/financial documents never leave the machine. PyPDF chunking → MiniLM embeddings → FAISS vector store → LangChain QA chain, served on a Chainlit chat UI, all on consumer CPU hardware.
A three-model multimodal pipeline: BLIP looks at an image and captions it, GPT-2 expands the caption into a children's story, and a VITS voice model narrates it to MP3. Built for accessibility — visually impaired kids can hear the story inside any illustration.
A CNN trained on FER-2013 that watches a live webcam feed and classifies faces into 7 emotions in real time. Haar Cascade face detection feeds 48×48 crops into a hand-built 5-block Keras network — useful anywhere from driver-drowsiness to retail sentiment.
Unsupervised learning over Spotify's audio features — danceability, valence, tempo, energy — clustering tracks into acoustic buckets and recommending neighbours by coordinate distance. Playlist curation without a single manual tag.
A lightweight desktop browser on PyQt5's Chromium engine — navigation, address bar, full HTML5/CSS/JS rendering. The kind of single-purpose, tracker-free container you'd want for a kiosk or an internal tool.
An offline-first Tkinter task manager with flat-file persistence — no cloud, no latency, no credentials to leak. Simple software that respects the user's disk and privacy.
EDA over hundreds of Indian startup funding events (2015–2017): which cities own which sectors, where the big tickets flow by stage, and how shocks like demonetization bent the deal-volume curve. Pandas pipeline for cleaning messy currency strings and location aliases, Matplotlib/Seaborn for the story.