Ameya Kulkarni

       

currently: building in public, shipping tools that do exactly what the README says

• Founded Benchtop Labs, an open verification registry and quality benchmark for independent developer tools with an automated 6-point isolated sandbox test spec
• Shipped Cuttle (website), a 100% offline right-click file converter utility for Windows Explorer with sub-20ms startup and zero data uploads
• Shipped Talanton, a local-first Python library that sets automatic spending limits on AI API calls, live on PyPI
• Evaluated AI coding agent benchmark tasks for OpenAI & Anthropic models on Project Terminus-2nd-Edition by Snorkel AI at AirDawg Labs (Jun 2026 - Aug 2026)
• Maintaining Mustel, a Python CLI and MCP server, now past 5,500+ PyPI downloads
• Founded The Free University with 100,000+ lifetime visits and 1,000+ LinkedIn followers

Where useful open source lives. An automated verification registry for high-craft developer tools.

An open quality standard and automated registry that tests independent developer tools in isolated sandboxes before featuring them, replacing popularity contests with objective technical verification.

• Automated 6-point sandbox verification: checks architectural documentation, OSI license metadata, semantic releases, 365-day commit activity, clean ephemeral container installs, and green CI
• Strict binary evaluation: zero paid placements, no sponsored bypasses, and no partial credit
• Dynamic spec tag generation with auto-updating SVG badges and edge-cached JSON/Markdown metadata manifests for AI agents (/llms.txt and /api/projects)
• 5 verified founding tools (Rollit, Talanton, Mustel, Cuttle, Driftwood) with 22,000+ cumulative developer installs

# Inspect all verified tools via the open agent API
curl -s https://benchtoplabs.vercel.app/api/projects?format=md

The honest scale for AI spend, weighing the cost before the call, not after.

Local-first Python library that puts a real spending limit on AI API calls, checked before the request is sent, not logged after.

• Pre-flight token metrology using each provider's real tokenizer plus chat-template overhead accounting, exact cost computed before a call goes out across 60+ models on OpenAI, Anthropic, and open-weight providers
• BudgetGuard intercepts every call with hard and soft dollar limits, raising HardBudgetExceededError and halting the request in under 2ms if a threshold is breached
• One-line wrappers (wrap_openai) plus integrations for LangChain, LangGraph, FastAPI, and LiteLLM Gateway
• 100% local: every call logged to an embedded SQLite database, zero prompts or completions ever leave the machine
• Benchmarked at 24,586 calls/sec with zero lock errors across 25 concurrent threads, ~310 bytes stored per call

talanton cost "Analyze this" --model gpt-4o -o 500   # exact dollar cost before you send it

A second pair of eyes for your AI coding agent, one that never hallucinates a finding.

Non-AI static analysis layer for AI coding agents (Cursor, Claude Code, Windsurf). Deterministic scanning, no model calls, no hallucinated findings.

• Orchestrates three engines (Ruff, Bandit, pip-audit) plus 21 custom YAML pattern files covering subprocess, requests, flask, django, pickle, asyncio, and more
• Normalizes everything into a schema-versioned JSON report, with a precomputed agent_prompt field so an AI agent can act on it without parsing the raw output
• Ships an MCP server (mustel serve) exposing review, review_file, env, and check_package tools for direct IDE and agent integration
• Benchmarked at 100% recall across 14 planted bugs in 4 test projects, 0 false positives on a clean baseline
• Past 5,500+ PyPI downloads and counting

mustel review        # scan current directory, outputs JSON
mustel serve          # run as an MCP server for AI IDEs

Right-click any file in Windows Explorer, pick the format you want, and get a converted copy instantly.

100% offline native Windows desktop utility that converts documents, images, video, and audio directly from the File Explorer context menu without uploading bytes to third-party servers.

• Sub-20ms launcher communicates via Windows Named Pipes (\\.\pipe\cuttle-<user>) with a persistent background Qt worker daemon in the System Tray
• Sacred originals guarantee: source files are never altered, deleted, or overwritten; conflicting target names auto-increment collision-free
• 300ms debouncer bundles multi-file selections into batch conversion queues bounded to half available CPU cores
• Frameless floating progress card with live cancellation and interactive Windows toast notifications ([Open File], [Show in Folder])
• Multi-engine architecture across PyMuPDF, pdf2docx, Pandoc, Pillow, LibreOffice, and FFmpeg with clear quality guarantees (Lossless, Lossy, Best Effort)

# One-click portable installer
curl -sLO https://github.com/Ameya79/cuttle/releases/latest/download/Cuttle-Setup-x64.exe

1,000 possible futures for a stock price, computed before your coffee finishes brewing.

Stateless Monte Carlo simulation API for stock price paths, built on Geometric Brownian Motion (GBM).

• FastAPI backend, fully stateless, no auth keys required, single POST /v1/simulate endpoint
• Vectorized NumPy computation returns 1,000 simulated paths in milliseconds
• Returns p10/p50/p90 percentile price envelopes along with annualized volatility and probability-of-profit
• Rate limited at both the Nginx layer (20 r/s, burst 30) and the app layer (100 requests per 5s per IP)
• Full stack runs via Docker Compose: FastAPI backend, Next.js frontend, Nginx reverse proxy
• 700+ simulations run to date

Rolling stats for NumPy arrays, without dragging pandas along for the ride.

Rolling window statistics for NumPy arrays, without pulling in pandas.

• mean, std, sum, min, max, zscore, normalize, apply, all under one consistent function signature
• Uses numpy.lib.stride_tricks.as_strided for zero-copy windowing, and locks the returned views read-only to avoid segfault-prone manual stride math
• Supports min_periods to mask incomplete windows instead of failing on them
• 1,500+ PyPI downloads, CI on every push via GitHub Actions

Compresses your images, then forgets it ever met them.

Flask app for image compression, resizing, and image-to-PDF merging, all done in memory.

• BytesIO-based processing, no files touch disk beyond the request lifecycle
• Resize by percentage, adjustable JPEG/PNG compression quality
• Merges multiple uploaded images into a single PDF
• Files auto-delete 5 seconds after generation, no logging, no stored metadata

Learning, minus the paywall and the parts nobody asked for.

A free course and certification aggregator I founded and grew from scratch. It manually vets free courses and certifications from providers like Google, Harvard, IBM, Cisco, and freeCodeCamp, and runs interactive in-browser learning spaces (including a Python zero-to-production track with live code execution) plus instant Python and web compilers, so people can learn without wading through paywalled content.

• 100,000+ lifetime visits, 1,000+ LinkedIn followers
• 260+ editorially vetted courses across 90+ institution hubs
• 4-Axis Editorial Audit Rubric & technical research engine The Print
• Built and SEO-optimized from scratch with SSG across 185 pre-rendered pages, zero paid promotion

AI Systems Evaluation Intern · AirDawg Labs  Jun 2026 - Aug 2026

Worked on Project Terminus-2nd-Edition by Snorkel AI. Evaluated AI coding agent benchmark tasks for instruction clarity, test alignment, rubric quality, metadata correctness, Docker/base image compliance, and solution leakage risks. Evaluated full task submissions by inspecting task environments, verifier reports, Dockerfiles, task.toml, and oracle solutions before making Accept / Needs Revision decisions. Also contributed as a task submitter before moving to the reviewer side. Ubuntu / WSL, Docker, Python.

Research Analyst → Data Analyst Intern · CoreLayer Labs (Review.AI)  Oct 2025 - Feb 2026

Built an LLM-powered product-page scraper cutting manual review time by 75%. Built a real-time Streamlit dashboard integrated with Google Sheets for team workload tracking. Researched and documented 50+ AI tools.

• Problem Setter · Shastra Coding Club, TCET. Designs and validates test cases for T&P coding assessments
• Mumbai Hacks 2025. Built LifeLink (AI organ donation platform), qualified for Round 2 offline of India's largest agentic AI hackathon (HealthTech track)
• CGPA 9.21 / 10 · B.E. Information Technology, Mumbai University

Mumbai · open to opportunities