currently: building in public, shipping tools that do exactly what the README says
- 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
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
BudgetGuardintercepts every call with hard and soft dollar limits, raisingHardBudgetExceededErrorand 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 itA 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_promptfield so an AI agent can act on it without parsing the raw output - Ships an MCP server (
mustel serve) exposingreview,review_file,env, andcheck_packagetools 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 IDEs1,000 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/simulateendpoint - 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_stridedfor zero-copy windowing, and locks the returned views read-only to avoid segfault-prone manual stride math - Supports
min_periodsto 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

