Projects — Jineshwar Nariani
Grounded AI products, data tools for vision and world models, and hackathon builds — 10+ hackathons so far, several of them winners.
ClaimCheck
Fact-Checks Claims Against a Product’s Own Docs · Live App
Paste a launch post, landing-page copy or an answer an AI assistant gave about a developer tool; ClaimCheck splits it into claims and gives each one a verdict — supported, contradicted, not in docs or unverifiable — with the doc passage behind it. The rule the whole app is built on: a verdict stands only if its quote appears word for word in the docs. The model proposes verdicts and quotes; deterministic code checks every quote and downgrades any it can’t find. Claude Haiku splits the text, hybrid search finds passages, Claude Sonnet judges, verdicts stream in live, a teammate reviews them, and a weekly job re-reads cited pages and flags evidence that has disappeared. Across 45 judged claims in three test runs it produced zero false “supported” verdicts, and every model choice was made on measured cost.
TypeScript, React, Claude Sonnet, Claude Haiku, RAG, Grounding, Firecrawl, Background Jobs, RBAC, Evals, Vitest, Playwright
Recruiter Search
Full-Stack AI Talent Search · Next.js, Supabase, Stripe
Recruiters describe who they want in plain English; Claude Haiku turns the brief into structured filters (role, seniority, city, skills, employers, school, years, open to work) and one static SQL query finds the people. The model never sees candidates, and it can’t invent filters: every field is an enum of values that exist in the data, anything the data can’t check is returned as “unsupported” rather than guessed, and the server re-validates and clamps every output. Searches are charged atomically under a row lock so concurrent requests can’t overspend a free plan, results are snapshotted for fast stable paging (~40 ms), and editing a filter re-runs the search without the model. Row-level security isolates each company, and Stripe webhooks are signature-checked, deduplicated and applied in order. About $0.0002 per search with prompt caching.
Next.js, TypeScript, Supabase, PostgreSQL, Row-Level Security, Stripe, Claude Haiku, Structured Outputs, Zod, Prompt Caching
Kausal
Action–Video Causality for World-Model Training Data
A world model learns what happens next from gameplay recordings — but only if the logged actions actually cause what’s on screen. Kausal measures that for any game: it reads the camera’s response with phase correlation, finds the lag, gain and coherence between mouse input and motion, and tests it with a permutation test and false-discovery control, then drops or repairs desynced windows. Built as an operator for the Kuration SDK and run on 11.8 hours of Open P2P gameplay plus 30 CS:GO recordings. Where the SDK’s built-in check flagged 0–30% of desynced clips, Kausal flagged 100% with the exact lag and no false alarms; it found Quake recordings logged a frame late, and filtering a contaminated corpus improved a downstream inverse-dynamics model (62.0% vs 59.9%, p ≈ 0.009).
Python, Computer Vision, World Models, Phase Correlation, Permutation Testing, FDR Control, PyTorch, FFmpeg, Data Curation
Model’s-Eye Transcoding
When Is Lossy Video Compression Worth It for a VLM?
A vision-language model never sees your video file — it sees 32×32-pixel, two-frame tubelets sampled at about 2 fps. So the right format is the model’s own input grid, and the real cost of a large file is decode work the sampler throws away. Re-encoding to Qwen3-VL’s grid (672×384, 2 fps, H.264 CRF 33) makes files 18% of the original size and 6.4× cheaper to read, with no significant accuracy loss across 270 VideoMME questions (2,160 VLM answers, cluster bootstrap and McNemar tests against a timestamp-jitter noise floor). It pays for itself after 1.6 reads; the usual “just use HEVC” answer makes every read slower. Includes a metadata-only decision rule that triages a corpus over HTTP range requests without downloading it.
Python, VLMs, Qwen3-VL, MLX, FFmpeg, H.264 / HEVC, Statistical Testing, Cost Modelling
Clause
AI Legal-Enforcement Platform · Best Use of AI Winner
Upload a lease, medical bill or contract; Clause finds violations, calculates damages and drafts demand letters citing Massachusetts statutes. It drops any claim it can’t trace to a retrieved statute rather than let the model guess, keeps personal data local behind a leak-free redaction pipeline, flags illegal clauses live on Airbnb via a Chrome extension, and timestamps evidence on Solana.
Snowflake Cortex RAG, Gemini, FastAPI, MCP, Solana
Maple AI
Emotionally Intelligent Learning Companion · Best AI Project with Databricks Winner
Breaks hard topics into steps and notices when a student is frustrated or confused — Hume EVI reads emotion in voice — then adapts its guidance. A Databricks RAG stack (DBRX, LanceDB, LlamaIndex) keeps answers grounded.
Hume EVI, Databricks, LangChain, Next.js, Flask
EchoHands
Real-Time ASL Learning · Two Awards at HackHer
Duolingo doesn’t teach American Sign Language, so I built something that does: MediaPipe pose estimation gives instant feedback on each sign, with custom PyTorch recognition models and AI tutoring. Drew interest from venture investors.
MediaPipe, PyTorch, TensorFlow, RAG, Next.js
GritHub
GenAI Personal Sports Coach
A 10-step training regime built on Amazon Bedrock, form feedback from a vision-language model (UForm-Gen2 on Cloudflare) analysing training video, and emotionally aware 1:1 voice coaching through Hume EVI 2 — with progress tracked in MongoDB Atlas.
Amazon Bedrock, VLM, Hume EVI 2, FastAPI, MongoDB
Artsee
Agentic Creative Collaboration
Each artist gets an autonomous Fetch.ai uAgent that holds their creative profile and, by design, finds collaborators and negotiates terms on their behalf. Computer vision and vision-language models (VLMs) read the artwork itself to understand each artist’s style; Claude handles agent-to-agent conversation and Groq provides fast inference.
Fetch.ai uAgents, Agentverse, Computer Vision, VLMs, Claude, Groq
AISH
AI-Powered Terminal · CUhackit 2025
A terminal that thinks, speaks and automates: voice interaction, web search through Perplexity, AI reasoning with DeepSeek and Qwen, and one-command GPU environment setup (CUDA, cuDNN).
LangChain, Hume AI, Perplexity, CUDA, Flask
LLM 10-K Analysis
Automated SEC Filing Analysis
Reading 10-K filings reveals a company’s risks, financial health and strategic direction — but doing it by hand is slow. This pipeline downloads the 10-Ks for selected companies from 1995–2023 with sec-edgar-downloader (handling companies with fewer filings), cleans and chunks the text, and uses GPT to extract trends in risk factors, management discussion and sentiment. The results are shown as word clouds, trend graphs and sentiment charts in an interactive web app: a React/Vue front end over a Flask/FastAPI back end, with Plotly and D3 visualisations.
Python, sec-edgar-downloader, OpenAI GPT, Chunking, Flask, FastAPI, React, Plotly, D3
Workmark
Co-Founder · See the Founder Feather
A platform where CS students build verified work experience through collaborative GitHub projects and micro-internships with startups, SMBs and nonprofits.
Founder