Ishir Vasavada

Product leader · Platforms, data & AI · Techstars founder

I build the products other teams build on: platforms, data and AI.

I've spent 6+ years on enterprise platform and data products used by thousands of companies, and today I own a $20MM ARR developer platform. I started in finance, worked as a technology consultant and co-founded a Techstars-backed startup, so I'm as comfortable with the P&L and the architecture as I am with customers.

Try the products ↓ LinkedIn ↗ Resume (PDF)

~/agents · demo runsrunning

▸Turn a billing policy into a rulerules agent
│read data model → drafted rulev1
│tested 12 cases → 1 failed → fixed itv2
└12 of 12 pass · waiting for approval9.8s
▸Is this pizza box recyclable?vision model
│photo + program rules → model88% sure
└lid yes · greasy base no · +10 pts1.2s
▸Find "what do you see" in a videoRecognito
│searched every spoken word48 words
└2 hits · 0:09.8 and 0:20.60.02s

$20MM

ARR developer platform I own, including the agentic rules engine: wide adoption, 3× faster customer setup.

Millions

of records in the metadata repository I built, the data model behind every product team's objects.

$15MM

in deals opened by a new in-country data-center region, launched with Security and Infrastructure.

3,500+

enterprise customers on the Responsible AI and data governance products I launched.

Sheet 01 Work

Five products you can try

An enterprise platform, a startup, a side project and two agent systems for a construction business I support on technology and product. Each one runs here in the browser. Pick one.

Agentic rules engine

The pattern behind the agentic rules engine I launched and scaled across enterprise customers: describe a billing policy, and an agent drafts the rule, tests it against known cases, fixes what fails, and waits for an admin to approve. Pick a policy.

MCP contextevalsguardrails

Who it's for: Billing admins and implementation teams at enterprise subscription companies
Market: Enterprise subscription billing (B2B SaaS)
Problem: Every enterprise bills a little differently. Custom pricing and decision logic lived in custom code, services projects or white-labeled third-party decisioning software, which slowed every implementation and pushed key logic outside the platform.

12 / 12 cases passedready for approval

Agent trace

An illustration of the pattern, written for this site. Not my employer's code, data or UI.

Ecodrop Technologies

The company I co-founded (it launched as Conserve). Snap a photo of an item, learn whether it's recyclable, and earn points toward partner rewards. A campus leaderboard and a live feed turned recycling into a game. Pick an item to scan and compare the computer-vision version we shipped with a multimodal LLM rebuild.

computer visionReact NativeTechstars · AWSgamification

Who it's for: University students and staff, plus local merchants who fund the rewards
Market: University recycling and sustainability programs
Problem: People want to recycle but don't know what's accepted, and nothing rewards doing it right. Wrong items contaminate whole loads and cost programs money.

10,000+users

15+merchant partners

4university customers

$8.5KMRR

$200K+raised, Techstars and AWS

Take a picture of your materials

Rewards · 140 points

The computer-vision trace follows the app's own matching code: labels from a cloud vision API, split into words, checked against a materials list, then a product list. Vision labels and LLM outputs shown are representative examples. Reward ladder and 10 points per item are from the original partner deck.

Recognito

Search inside a video for the moment a word is spoken. The web app transcribed uploaded video with cloud machine-learning models built on recurrent neural networks, and a Chrome extension prototype brought search to any page's video player. Search the clip below, then jump to any hit.

speech alignmentsearchbrowser extension

Who it's for: Journalists, legal teams, researchers and everyday viewers of long video
Market: Media houses, bar associations, public-affairs archives and video platforms
Problem: Finding one moment in a long hearing, lecture or interview means dragging the time bar or digging through comments. Search should work on what was said.

TRY

Transcript · click any word

Audio16 kHz mono
Aligneach word to the waveform
Indexwords
Searchword, prefix, phrase

Word times come from forced alignment of the clip's captions with CMU PocketSphinx, accurate to about a tenth of a second. Clip: Elephants Dream (2006), Blender Foundation, CC BY 2.5.

How the first version worked

v0 · Chrome extension

Recreated from the original popup.html and popup.js.

What was real

// video_control.js, the v0 content script
chrome.runtime.onMessage.addListener(
  function(stamp) {
    var video = document.querySelector("video");
    video.currentTime = stamp;
    video.pause();
  }
);

The popup and player control worked on any page with a video. Transcription ran in the web app; the extension's search results were still placeholders that returned the same three timestamps for every query.

What this demo adds to the extension

Search: ~~3 fixed timestamps~~ every spoken word, indexed
Precision: ~~to the second~~ to about 0.1 s, per word
Queries: ~~one word~~ word, prefix or phrase
Results: ~~a list~~ marks on the timeline and a live transcript

Blueprint vision agent

Built for a residential construction business I support on technology and product. A multimodal agent reads an architect's drawing and returns structured data: every room, its dimensions and square footage, ready for an estimate. It flags what it isn't sure about instead of guessing. Click any room, or reclassify one and watch the totals follow.

visiondocument extractionhuman review

Who it's for: Estimators and project managers at small residential builders
Market: Residential construction
Problem: Every estimate and material order starts with measuring each room off the drawings by hand. It is slow, and one misread dimension flows into the whole budget.

Agent trace

1,824

Heated sf

684

Unheated

2,508

Gross

Cost-control agent crew

Built for the same construction business, where I lead the technology and product side. Type a cost the way you'd text it. An orchestrator hands it to a small crew of agents that file it, check for duplicate payments and test it against the budget. Clean entries post on their own. Anything risky waits for a person.

multi-agentorchestrationhuman approval

Who it's for: Owners and field crews at small builders, who are not spreadsheet people
Market: Small construction businesses
Problem: Costs get paid on site and never reach the spreadsheet. The budget was always behind, so overruns showed up late.

9:41Sample project

Job Cost

Two-lot build · sample data

Budget used

Latest

Agent crew · runs on every entry

Try one

The shipped app replaced a spreadsheet that kept falling behind: crews log costs from the field, so the budget stays current. It turned a texted cost into a ledger row in one step. This demo shows the multi-agent version, with sample data.

Also built: a code-to-spec tool that turns NFRC, NAFS and IRC glazing and egress requirements into supplier specs, used to qualify overseas window and door manufacturers. Direct sourcing saved $75K+.

Sheet 02 Method

How I build with agents

The same loop for an enterprise platform or a field tool. The model is the cheap part. The context, the evals and the approval step are the product.

A few things I've learned building agents

One agent, until it isn't enough

Most problems need one good agent, not five. I split the work only when the pieces can truly run side by side, like the cost-control crew.

Let the platform run it

Managed agent runtimes now handle sandboxes, memory and long sessions. Good. My time goes to what the agent can touch and when a person steps in.

Evals come first

If I can't write the test cases, I don't understand the feature yet. Every miss in production becomes a new case.

Context beats prompts

Agents fail more from what they can't see than from bad wording. Few tools, clean context, and MCP for the plumbing.

Trust is earned

I start with a person approving everything and loosen it as the evals hold up. Anything you can't undo keeps a human.

Every task has a price

Finance taught me to price things, and I still do. Each task gets a cost and latency budget, and the cheaper model goes first.

Sheet 03 Path

Finance, consulting, product

I started in finance, moved into technology consulting, then into product management. Co-founding a Techstars-backed startup taught me to build from zero, and I bring that to enterprise products.

201820192020202120222023202420252026

Enterprise

Founder

Range

Product strategy, discovery and go-to-market

Platform, API and data architecture

Pricing, packaging and unit economics

Hands-on prototyping, from mobile apps to agents

AI credentials

AWS Certified AI Practitioner

Azure AI Fundamentals (AI-900)

Agentic AI (Andrew Ng)

Google Cloud Generative AI Leader

AI Product Management Certification, Product School

Also

U.S. Patent Pending 63/502,475

Certified Scrum Product Owner

Product Analytics, Product School

SAFe 5 Agile

Sheet 04 Contact

Let's compare notes.

I'm always glad to talk product, platforms and where AI is heading. Email is the fastest way to reach me.

ishirv00@gmail.com LinkedIn ↗ Resume (PDF)