Building applied-AI products? Let's talk.
Email is the fastest way to reach me.
Jaskaran Singh Kalra
Product Manager · Applied AI · San Francisco
Nine years taking AI products from 0 to 1. I write the evals that decide when a product is ready to ship, and I drive delivery across engineering, data science, BD, and operations.
Experience
Analog Devices (ADI)
Data Product Manager, Central Data Office
Oct 2024 to present
Uber
Product Manager / Product Operations Manager, Marketplace Products
Oct 2021 to Apr 2024
Tyrula
Product Manager, Enterprise SaaS
Mar 2017 to Sep 2021
Duke University
M.Eng. Engineering Management (Product Management)
Work
Internal product names, data, and screenshots are left out.
ADI · LLM evals · Release gating
The eval suite that decides when an enterprise AI agent ships
Sole PM and engineer. The agent answers plain-English questions about company revenue. I built 40 business-critical test questions with frozen correct answers, checked against approved finance reports, plus an LLM-as-judge: a second model that grades every answer correct, partially correct, or hallucinated.
It replaced 120 engineer-hours of manual testing per release. It blocked 2 production releases until defects were fixed, and both then shipped with zero regressions. The same results justified moving the agent from Opus to Haiku, a smaller model, cutting cost and latency.
ADI · LLM product · Adoption
A second brain on Claude, built to replace the wiki
Built and shipped a knowledge product on Claude that replaces knowledge bases like Confluence. It breaks meeting transcripts and work documents into linked pages anyone can ask questions of.
40+ engineers, PMs, and leaders adopted it within 3 weeks. Each now resolves 5 to 10 questions a week that used to wait on one overloaded expert.
Tyrula · Document AI · 0 to 1 SaaS
A commercial document-AI product, still sold today
Turned a back-end data entry tool into a commercial AI document-processing SaaS, owning product strategy, delivery, and customer-facing quality end to end. It cut manual effort 90% across ~4,000 forms scanned daily, and the legacy tool was retired with no customer churn.
I designed the trust layer that got non-technical users to accept the model's output: low-confidence extractions go to a human for review, with field-level accuracy targets and an audit trail. 98% CSAT across enterprise accounts.
ADI · ML · Anomaly detection
Grey-market detection at invoicing, not months later
Defined, modeled, and shipped an ML anomaly-detection product across 2,000+ product lines. It catches grey-market activity: products resold outside approved channels.
Detection now happens at the time of invoicing, instead of ~6 months after the sale. I built the detection model and the analyst-facing interface myself.
Uber · Roadmap · ML systems
A 3-year roadmap across 40+ stakeholders that won $5M in funding
Authored the CPO-approved 3-year product vision and roadmap across 40+ stakeholders. ~90% of Year 1 was delivered, and it won $5M in incremental funding for the team's annual runway. Held go/no-go authority over third-party data vendors, leading technical negotiations alongside BD.
Owned the ML system that estimated competitor market share. It powered a merchant pricing product and investor reporting read by senior leadership up to the CEO. I lifted its uptime from 65% to 95%.
Built on my own time
RAG · Retrieval evals
Local search over my own notes, measured instead of eyeballed
A RAG system (the model looks up my documents before it answers) that takes plain-English questions about my entire Notion workspace. It runs fully local: Ollama for the models and embeddings, Chroma as the vector store, Streamlit for the interface.
I built a golden question set and scored retrieval with Precision@5 (how many of the top 5 results are actually relevant) to compare chunking strategies and overlap settings.
Point of view
Most AI features fail because nobody defined what good output looks like.
That's a product spec problem, not a model problem. The PM's job is to write the eval before the engineer writes the prompt.
The best AI PMs I know are the ones who've broken their own automations.
Shipping a personal AI tool teaches you more about failure modes, latency tolerance, and user trust than any case study.
AI is making product strategy harder, not easier.
When any feature is buildable in a week, the constraint shifts entirely to judgment: what should we build? That's a harder problem with more surface area for bad decisions.
Email is the fastest way to reach me.