Seeking AI product management internships (Summer 2027) and 2027 new-grad roles
AI product · agent systems · San Francisco, CA · Open to relocate
Jaideep Aher
I decide what an AI agent should do for its users, then build the first version to see if I was right.
Duke MEng AI ’27 · Fuqua AI Product Management · YC Startup School ’26 · 2 ventures co-founded
For four years my job has been deciding what to build and then building it. That covers two ventures in fleet safety and agri-tech, agent workflows at ProcessMaker, and ThriftySpark, a thrift-store SaaS with paying stores that I re-aimed in a day with twelve agent sessions. The question I keep coming back to is when an agent should act and when it should hand the step to a person.
Projects
Selected work
Agent systems and the products around them first, then three Android apps for a quieter phone, then ventures with users and revenue, then applied AI builds. Each entry says who had the problem, what shipped and what changed as a result.
Agent systems and the products around them
Work where I decided what an agent should do and when it should hand off to a person, then built it. Each has a write-up or source code to read.
ThriftySpark · founder, PM and orchestrator
AI for thrift stores, with stores selling on it. 25,000 US thrift stores are not searchable online because listing costs more labour than a $3 shirt earns; ThriftySpark takes a photo to a full listing, price, SKU and inventory record in seconds, and runs the back room and the till on the same stock count. The write-up covers the problem, the economics, the market, the competition, the business model and how I run AI agents as the team.
- Founder
- Multi-agent orchestration
- Market research
- SaaS
- POS
Mira · self-auditing clinical intake
Chronic illness patients wait months for a short visit and arrive with scattered notes. Mira turns those notes into a doctor-ready timeline, then a second pass audits the first for missed negations, invented symptoms and out-of-order events before anything is shown. The bet was that a clinician will only trust an agent that checks its own work.
- Gemini
- DistilBERT
- Self-audit loop
- Next.js
Course Co-Pilot · Duke's Spring 2027 courses, filterable
Transfer credit evaluation takes a registrar days per student. I scoped a prototype as an agent pipeline: one agent reads the transcript, one checks the receiving catalog, a third grades its own confidence and routes uncertain cases to a human. The catalog layer is now live: every open Duke Spring 2027 course, tagged by skill and AI resilience, with unsure tags sent to review instead of the page.
- GPT-4o
- Multi-agent
- Human in the loop
- Live catalog
RoboTalk · constrained voice commands for robotaxis
A robotaxi cannot obey every voice around it. RoboTalk turns typed or spoken requests into a constrained command set, checks who is asking and what they are allowed to do, and shows the decision in a live 3D scene. Underneath, it is a permissions model, and the agent in question weighs two tonnes.
- NLP
- Safety-critical AI
- Permissions
- Three.js
Android apps for a quieter phone
The launcher locks the feeds, Drab Tube gives boredom one narrow outlet, and Jarvis rings when a plan slips. Each has a case study and a test build.
Test-signed build. Android 9 or later. Download it and open the file to install; Android will warn about an unknown source.
Drab Launcher
Phones are built to be interesting. This Android home screen puts a growing wait in front of every app you mark and keeps the next task one tap away. Loosening any rule waits on a timer.
- Android
- Kotlin
- Accessibility service
- Behaviour design
Drab Tube
When the launcher locks YouTube, the urge to watch still needs somewhere to go. Drab Tube is the only app left that plays video. It holds ten study series with no feed and no search, and the next ten open after seven are done.
- Android
- React
- TypeScript
- Behaviour design
Jarvis
A reminder is easy to swipe away. A coach server decides when to call, and this app rings like a real phone call, joins a voice call on Answer and reports what happened.
- Android
- Kotlin
- Firebase Cloud Messaging
- WebRTC
Without a coach server to pair with, the app only shows its pairing screen.
Ventures with users and revenue
Two companies I co-founded and took from idea to ₹7.2M (about $86K) combined ARR, and the farmer platform that grew out of one of them.
Fleet Kawach · co-founder, product
Fleet operators lose money and lives when unsafe driving goes unchecked. I ran discovery with operators, chose in-cab coaching over after-the-fact reports because a warning at the wheel can change the drive while it is happening, and led the 0 to 1 build: risky-behaviour detection, drunk-driving prevention and real-time coaching.
- Computer vision
- Edge AI
- IoT
- B2B SaaS
Agrofix · co-founder
Farms over-spray chemicals, which wastes money and damages soil. I turned two granted patents from IIT Ropar and a $13,000 iHub AWaDH grant (chosen from 800+ applicants) into a targeted spraying product that treats only the areas that need it, and chose a hardware-plus-service model so smallholders pay by the season instead of buying a machine.
- Computer vision
- Robotics
- Agri-tech
- Hardware
Hundekari · app plus portal, about 8K farmers a month
Smallholder farmers kept their business on paper and got poor prices because they did not control their own records. I scoped an Android app around the three jobs they ran on paper, then a web portal with digital receipts, local rate history and a verified marketplace. The app has 5K+ downloads and, with the portal, reaches about 8K unique farmers a month.
- Android
- Marketplace
- ERP integration
- Agri-tech
Applied AI builds
Coursework, research and independent builds. Two run in your browser on this site.
Flow Matching Lab · runs in your browser
Flow matching powers current frontier image models but is usually explained in equations. Draw any shape and watch a hand-written neural net learn to generate it from noise, live in the browser.
- Flow matching
- Generative AI
- Backprop from scratch
- TypeScript
SignalForge · auditable AV scenario generation
AV safety teams need test scenarios they can trace to a regulation or an actual crash, and a prompt does not count. SignalForge expands safety cases into concrete driving conditions, renders synthetic lidar and radar, and shows the source behind every scenario. Built from what I saw certifying L3 at Magna.
- Autonomous vehicles
- Scenario generation
- LiDAR and radar
BinSnap
Recycling classifiers fail on the dark, blurry photos people actually take at the bin. BinSnap is trained for those photos, shows its confidence, and explains what it looked at.
- PyTorch
- MobileNetV3
- ONNX
- React
VoiceClone AAC
People with ALS can lose their own voice. An iOS app that clones a voice from a short sample and speaks any typed message back in it.
- iOS
- Swift
- Voice cloning
Also built
A browser LiDAR viewer for KITTI scans, a privacy-first driver monitoring system that sends signals instead of video, a child-safety message classifier that stays visible to the child, a pneumonia X-ray screening model, a side-by-side fine-tuning demo, a reef health scorer, and a reusable container system for campus dining.
- Computer vision
- NLP
- Edge privacy
- Full-stack
GitHub
Contributions
Public commit activity over the last year. @jaideep-aher
Rendered by githubchart from public contribution data; it may differ slightly from GitHub's own chart.
Contact
Hiring an AI PM who can also build the agents?
Seeking AI product management internships (Summer 2027) and 2027 new-grad roles. To see how I work, read the ThriftySpark write-up. It covers the problem, the economics, the market, the decisions, and how I run AI agents as the team. Email, LinkedIn or book a time at aher.dev/meet.
San Francisco now · Duke MEng through May 2027 · Open to relocate