[ok] Marketplace Negotiation Agent

Capture listing -> Vision model prices it -> Negotiates autonomously

A browser extension plus a local server drives an operator's own authenticated session on a large consumer marketplace: assesses item condition from photos, prices accordingly, then negotiates with sellers autonomously — with rate-limit detection and messaging guards built after real incidents, not speculative edge cases.

# runs continuously on a cloud instance, negotiating in parallel

[ok] Sports Capture Quality Control

Device upload -> Parameter + CNN checks -> AI fallback -> Annotated video

Two production AWS Lambdas validate every capture a device uploads: parameter checks, then CNN/ONNX visual verification (Bedrock Nemotron where no model was trained for the mode), then a rendered debug video for anyone reviewing the shot.

# 86% recall on range-mode ball-impact detection · ~16s avg render time

[ok] Launch Monitor ML Suite

Collect 1,218 exports -> Engineer features -> Train 3 models -> Ship to iOS

Scraped bulk sensor exports from a third-party sports launch-monitor portal, then trained neural nets that predict spin rate, spin axis, and club speed from the metrics that are easy to measure — shipped as CoreML models for on-device iOS inference.

# club speed error cut from 20.5 km/h (legacy) to 1.99 km/h

[ok] Tournament Data Agent

Discover pages -> AI extraction -> Read flyer images -> Merge & dedupe

An autonomous agent discovers charity sports-event pages across a tournament-management platform, decides for itself how far to navigate to reach rosters, extracts structured data with a vision-capable model, and reads promotional flyer images via OCR when the details were never real HTML text.

# 178 tournaments extracted end to end from unstructured event pages

[ok] Professional Network Automator

Browse leads -> Human-paced engagement -> Email alerting

Runs locally and on a cloud instance independently, working through a lead list on a professional network and engaging with the content worth engaging with — randomized human-like pacing, daily caps, and crash-alert email built in after the process silently went down more than once.

# 1,541 items tracked · 98.8% success rate

[ok] Web Scraping Fleet

Browser extension -> Local server -> Deduplicated dataset

Five independent scrapers, each a purpose-built Chrome extension paired with a local server: association-executive directories, a speaker marketplace, a member community feed, and a professional network — each with its own dedup and resume logic.

# 22,000+ profiles and posts collected across 5 platforms

[ok] Deep Listening

Surveys & uploads -> Multi-modal extraction -> Fine-tuned LLM -> Vector retrieval

A multi-modal information-extraction platform. Hugging Face models handled object detection, tag generation, and image description on images; translation and transcription on audio; both combined for video; summarization and keyword extraction on text. A fine-tuned LLaMA drove additional extraction and chat, backed by a vector database for retrieval.

# text, image, audio and video in one extraction pipeline

[ok] CipBot

WhatsApp -> Structured chat flow -> Client database -> Admin panel

An intelligent chatbot connecting Cipla's procurement team with its vendors, deployed on WhatsApp. A structured chat flow surfaced internal database records, with role-based options per user type and an admin panel for reviewing interactions and data traffic.

# built on Rasa, live on WhatsApp for a pharma procurement team

[ok] Trajectus IVR

Phone call -> Voice IVR -> Medical query handling

An IVR-based chatbot handling medical query calls, built on Google DialogFlow. Delivered as outsourced work for Computer Instruments (CI).

# voice-first, not text — a different failure surface entirely

[ok] Coglide — Cori & Gx

Inbound email -> Contact extraction -> Extension sync -> Marketing database

"Cori", an email service extracting contact details out of inbound mail, plus "Gx", a browser extension syncing data from diverse marketing websites into Coglide's own database. Also owned the data scraping and automation work across the product, and ran the research server that fed its marketing data.

# the earliest version of the scraping work that runs through everything since

[ok] Emotyx

Live CCTV -> Backend network -> Customer insights

Backend network structure for software extracting intelligent, actionable insights from real-time CCTV footage for customer understanding — built and deployed in Python, handling the managing and recording of the footage itself.

# real-time video in production, not batch analysis

[ok] Attendance Management System

Student records -> Python backend -> Image processing

Owned the entire back-end code structure for managing student and attendance records, and supported the AI development around its image processing.

# full backend ownership, end to end

[ok] Other Projects

A GPT-3.5 minutes-to-meeting generator that transcribed calls and wrote up the minutes. Shell tooling to verify API endpoint data and harden it against SQL injection, shipped as a Postman collection. A Selenium scraper collecting marketing data across multiple social networks. A back-end network for the Learning & Development team, and a scraping service for Talent Acquisition to surface new candidates. Plus the back-end and detection module for PPE (Personal Protective Equipment) kit detection.

# the long tail — smaller builds, same instinct