Projects

Wearable AI project · 2026

Meals that log themselves

Problem

Counting calories without ever taking a photo: a wearable sees a stream of images, and the wearer should do nothing.

Built

The whole flow from images to a few correct meal entries, with every estimate grounded in real food data instead of the model’s guess.

Result

~25% lower calorie error than the vision model alone, on a public benchmark of 150 dishes.

• Computer vision
• Multimodal AI
• Grounding
• Model evaluation
• Benchmarking

See it in the journey →

Forkast · Apr – Oct 2026

Live competitor prices for restaurant chains

Problem

Restaurant chains set prices from spreadsheets.

Built

Live competitor prices collected store by store from each brand’s own ordering site and from DoorDash, Uber Eats and Grubhub, across 140,000+ stores.

Result

Won Forkast’s first paying customer, a chain of 500+ restaurants, which used it to decide which competitors to focus on, and where, to bring in more foot traffic.

• Web scraping
• Data pipelines
• Python
• Pricing analytics
• Product leadership

See it in the journey →

Forkast · 2026

An agent that writes its own scrapers

Problem

Adding each new restaurant brand took weeks of engineering.

Built

An agent that finds a brand’s stores, writes its own scraper and checks its output against an answer key. AI-written code never reached production without a sandbox, a judge model and a human review.

Result

0.96 price agreement across 140,000+ stores, and new brands without weeks of engineering. Every agent was red-teamed before release.

• AI agents
• LLMs
• Code generation
• LLM-as-judge
• Sandboxing
• Red-teaming
• Human-in-the-loop
• Playwright
• curl_cffi
• Airflow

See it in the journey →

Forkast · 2026

Ottoscope

Problem

Creators sign brand deals without knowing what they keep or give away.

Built

A product I launched on my own: a creator uploads a brand-deal contract and learns what money they actually keep, what rights they gave away, and when to renegotiate. Every number it shows quotes the contract word for word.

Result

Launched as Forkast’s second product.

• Document AI
• LLMs
• Citations
• Product launch

Forkast · 2026

A disputes system for a buy-now-pay-later lender

Problem

A buy-now-pay-later lender handled every dispute by hand.

Built

A first version of their disputes pipeline, built inside the disputes team alongside the people doing the work: a state machine whose rules live in a table, so adding a merchant is a row, not code. Claude only advises, and anything it is unsure of goes to a person.

Result

Built and validated against the lender’s real backlog in a dry run, with every outgoing message checked for personal data.

• Workflow automation
• Fintech
• State machines
• Databricks
• Claude
• Human-in-the-loop
• Customer discovery

Forkast · Sep – Oct 2026

A two-minute pricing audit

Problem

A restaurant brand had no quick way to see how its prices compared with the competitors next door.

Built

A public audit: pick a brand and see its nearest store against four to six competitors on prices, promotions and delivery, in about two minutes. It reuses the core platform instead of copying it, caches results and tracks the AI cost of every lookup.

Result

Our lead-generation tool, and the pitch in every sales email.

• Lead generation
• LLM extraction
• Cost tracking
• Caching
• Python

Forkast · 2026

Running go-to-market

Problem

A startup with one engineer still needs someone to find and win customers.

Built

I ran sales calls and demos, built the pitch: price-position reports and per-chain proof-of-concept dashboards, and shaped pricing and positioning. Then I designed an outreach engine to replace my manual prospecting, where a person approves every email.

Result

Hired interns and a contractor along the way, running the whole loop from sourcing to onboarding.

• Go-to-market
• Sales
• Demos
• Hiring
• Product strategy
• Next.js
• TypeScript

Forkast · Oct 2026

A new market: truck-rental pricing

Problem

An enterprise truck-rental company wanted to see where its prices stood against a competitor.

Built

Neither site had a public API, so I captured how their quote pages talk to their servers and replayed those calls across hundreds of routes.

Result

Our first market outside restaurants, and a pitch built on real quotes.

• Reverse engineering
• Pricing analytics
• Python
• Enterprise sales

Adsgency AI · Mar – Apr 2026

One schema for every social platform

Problem

Every new client meant onboarding their ad data by hand.

Built

Pipelines that read a client’s raw TikTok, Instagram, Facebook and X data and turn it into one clean schema, with fallbacks for when the generative AI making ad decisions gets it wrong.

Result

Nobody has to onboard a client by hand.

• ETL
• Data pipelines
• Data modeling
• Generative AI
• LangGraph
• Fallbacks

Adsgency AI · Apr 2026

Making the agent pipelines safe to run

Problem

A worker that crashed mid-job could hold its lock and stall a client’s feed, and some API keys travelled without full TLS.

Built

Guaranteed lock release with a 10-minute expiry, and closed the TLS gaps around client API keys.

Result

A crashed worker no longer stalls client feeds, and client keys stay protected.

• Reliability
• Distributed locks
• Security
• TLS
• Python

Asante · Dec 2025 – Mar 2026

GoLocal Maps

Problem

Businesses doing good in their community had nowhere to show it.

Built

The first version of golocalmaps.com, on my own: a public map of local businesses and nonprofits, including businesses that hadn’t signed up yet, placed by geocoding their addresses.

Result

Live at golocalmaps.com on AWS, with analytics to see how people use it.

• React
• Leaflet
• gRPC-web
• AWS
• Geocoding
• Full-stack

Asante · 2025

Data on 1.8 million nonprofits

Problem

The matching product needed real data on nonprofits, not a hand-made list.

Built

A pipeline that pulls nonprofits’ IRS filings through the ProPublica API in parallel, resumes where it left off, and maps the IRS categories onto ours.

Result

1.8M+ nonprofit records enriched with revenue, size and category.

• Data pipelines
• Python
• APIs
• Data cleaning
• PostgreSQL

Asante · 2025

Outreach agent

Problem

The founder was drowning in prospecting.

Built

An agent I proposed and built: it finds prospects, scores fit, writes personal emails and keeps HubSpot up to date.

Result

Gave the founder back ~90% of the time spent prospecting.

• AI agents
• LLMs
• HubSpot
• CRM integration
• Sales automation

See it in the journey →

Asante · 2025

Inbound reply agent

Problem

Inbound leads needed fast, accurate answers.

Built

A reply agent grounded in the company’s own knowledge, passing only the valuable leads to a person.

Result

~80% response rate on inbound leads.

• RAG
• LLMs
• AI agents
• Lead qualification

See it in the journey →

Asante · 2025

Nonprofit and business matching

Problem

Matching nonprofits with businesses, with no data to learn from.

Built

The matching product from nothing: I designed the survey that created the data, labelled it, trained a Random Forest on SageMaker and served semantic search over gRPC.

Result

Results served in under 200 ms.

• Machine learning
• Random Forest
• AWS SageMaker
• gRPC
• Recommendation systems
• Data labeling
• Model serving
• MLOps
• Backend

Best Use of Auth0 for AI Agents

Deep Agents Hackathon, RSAC 2026 · Mar 2026

Penny, an AI financial mentor for kids

Problem

Kids get money advice from nobody they can talk to, and parents want a say.

Built

Most of a video mentor, with Claude speaking through a Tavus avatar, that talks a kid through investment choices. Agents score the risk, and every recommendation goes to a parent for approval, with Auth0 deciding who can do what.

Result

Won Best Use of Auth0 for AI Agents, in a team of three.

• AI agents
• LangGraph
• Claude
• Auth0
• Human-in-the-loop
• Video AI
• FastAPI
• React

Code on GitHub →

Hackathon · Oct 2026

data-here: who to meet at an event, and why

Problem

A sales team at an event gets a list of hundreds of attendees and no idea who matters.

Built

An agent that learns what the seller sells, sorts every attendee’s company, pulls LinkedIn signals and ranks people over a knowledge graph that explains each pick. It runs as an MCP server, and no email goes out until a person approves it.

Result

A ranked list where every pick comes with its reasons.

• AI agents
• MCP
• Knowledge graphs
• FalkorDB
• OpenAI
• Bright Data
• Docker

Code on GitHub →

MeetStream AI hackathon · 2026

Sales intelligence inside Google Meet

Problem

Salespeople walk into calls knowing little about the person on the other side.

Built

A multi-agent system that researches the company and the people, a dashboard for preparing the meeting, and a Chrome extension that shows live insights during the Google Meet call. LinkedIn data comes through Apify, with enterprise SSO.

Result

A working end-to-end demo: research before the call, insights during it.

• Multi-agent systems
• MCP
• OpenAI GPT-4
• FastAPI
• React
• Chrome extension
• WebSockets
• SSO

Code on GitHub →

Hackathon · 2026

AutoOps: covering a shift when someone calls in sick

Problem

When an employee calls in sick, a manager scrambles to find a replacement.

Built

An agent that picks the best replacement from availability, reliability and hours already worked, and explains its reasoning with a confidence score.

Result

A replacement and the reasons for it, from one API call.

• AI agents
• Claude
• Retool Workflows
• FastAPI
• Supabase
• Explainable AI

Code on GitHub →

Hackathon · 2026

AI Virtual Tour

Problem

Apartment hunting means hours of clicking through listings and dense text.

Built

A voice assistant: say a property’s name and it finds the listing, extracts prices, floor plans and amenities, writes a summary with pros and cons, answers follow-up questions, and shows the right photo as you talk (“show me the kitchen”).

Result

Apartment search as a conversation, hands-free.

• Voice AI
• Web automation
• Tavily
• Multimodal AI
• Python

Code on GitHub →

Side project · 2026

Evaluation harness for pharma voice agents

Problem

Voice agents in a regulated field need to be tested before anyone trusts them.

Built

A test harness that runs a voice agent through scripted scenarios, scores each run, and rolls the results up into a dashboard with the pass rate, the top issues and a drill-down on every failure.

Result

One run covers every scenario, with the failures ready to read.

• LLM evaluation
• Voice AI
• Test automation
• Python

Code on GitHub →

Side project, in progress · 2026

Venue research agent

Problem

Finding and contacting the right venue takes hours of searching, reading reviews and writing emails.

Built

An agent that searches for venues, scrapes their sites, analyzes reviews, ranks them on several criteria and drafts personal outreach, with hallucination checks and a confidence score and source for every claim.

Result

Research and outreach in one run, with every claim traceable.

• AI agents
• Tool calling
• Hallucination detection
• FastAPI
• BeautifulSoup
• Hugging Face

Code on GitHub →

Side project · 2025

Symptom-to-disease assistant

Problem

People describe symptoms in their own words, not in a doctor’s terms.

Built

A triage-style assistant that matches free-text symptoms to known ones with semantic embeddings, suggests related symptoms to confirm, and ranks likely diseases.

Result

A FastAPI service with a Streamlit front end.

• NLP
• Embeddings
• sentence-transformers
• FastAPI
• Streamlit

Code on GitHub →

IIIT Kottayam, with a rubber-parts factory · 2022 – 2023

Vision model that inspects rubber washers

Problem

Workers checked every washer by stretching it by hand.

Built

A YOLOv5 model, trained on a dataset I curated, that spots bad washers on its own, deployed on GCP with a loop for workers’ feedback.

Result

91% mAP@0.5, 60% less inspection effort and 35% fewer errors.

• Computer vision
• Object detection
• YOLOv5
• PyTorch
• GCP
• Model training
• Research

See it in the journey →

Sold to Trell · 2020

Vendor spreadsheet automation

Problem

My sister typed delivery spreadsheets from many vendors into her company’s system, eight hours a day.

Built

A Python package that did the job for her. I showed it to her manager, and the company bought it.

Result

Eight hours a day became ten minutes, and missing deliveries dropped to almost zero.

• Python
• Workflow automation
• Data cleaning

See it in the journey →