DOC NO. DB-2026 / PORTFOLIO
BOOK A CALL ↗
STATUS: BUILDING
DEVANSHU BRAHMBHATT — TECHNICAL AI PM · BUILDER · RESEARCHER · FOUNDER
I AM TRYING TO GET VERY GOOD AT ONE THING →
Closing the gap.
BETWEEN
what frontier models can do
AND
what enterprises can actually deploy.
NOTE 01 — THE RECORD
▸ 0→1 AI PM for Skyflow Detect
▸ Former Rakuten software engineer
▸ Berkeley Lab researcher, first author
▸ Built two startups
SUBJECT: D. BRAHMBHATTSAN FRANCISCO, CALIFORNIA
FIELD NOTE 01
Hi. I want to spend my best years building products that matter, with people who care deeply about getting them right.
SHIPPED
Skyflow Detect
HYPOTHESIS → FORTUNE 500 PRODUCTION
ENGINEERED
Rakuten
SYNTHETIC TEST DATA → 10K+ HOURS SAVED
PUBLISHED
Berkeley Lab
FIRST AUTHOR · NATURE PORTFOLIO · ACM
FOUNDED
JobAuto + PaperTalk
7K+ USERS · $5K MRR
EXPERIENCE →
FIG. 02 — EXPERIENCE
Experience.
EXHIBIT A — SKYFLOW DETECT · 0→1 AI PRODUCT MANAGER
FEB 2024 — PRESENT · PALO ALTO, CA
$100M+ RAISED
BACKED BY KHOSLA · INSIGHT · FOUNDATION CAPITAL
FOUNDED 2019 BY EX-SALESFORCE & ORACLE EXECS
AI needs the data enterprises are least willing to expose.
THE PRODUCTDetect finds and tokenizes sensitive data across datasets, files, and APIs using Vault tokens.
AT RUNTIMEPolicy determines whether each value is returned as plaintext, masked, or redacted.
WHO IT SERVESAI, data, and security teams deploying models and agents on sensitive enterprise data.
MY ROLE0→1 AI PM. Turned customer needs, model constraints, and infrastructure limits into a product enterprises could deploy and trust.
WHAT CHANGEDTook Detect from hypothesis to production with Fortune 500 companies and leading AI labs.
THE TEAM — SKYFLOW HQ, PALO ALTO
JUDGING A HACKATHON — AWS BUILDER LOFT, SF
DATABRICKS DATA+AI SUMMIT
THE HARD PART
Detecting sensitive data is the demo. Everything after that is the product.
THIS IS WHERE MOST OF MY WORK LIVES: POLICY · RELIABILITY · MODEL QUALITY · SCALE · ECONOMICS.
{{ skyLabel }}
{{ skyArrow }}
01 · PRIVACY SEMANTICS
THE CHALLENGE
How can a model use sensitive context without receiving the raw values?
PRODUCT CALL
Governed tokenization instead of permanent redaction.
WHY IT MATTERED
Store the original value in the Vault, send tokens through model and agent workflows, and let policy control re-identification. Utility preserved without giving every actor unrestricted access.
02 · SYSTEM CONSISTENCY
THE CHALLENGE
How should privacy behave across databases, files, APIs, and agents?
PRODUCT CALL
One privacy model across structured and unstructured data.
WHY IT MATTERED
The same entity, token, and policy semantics behave consistently everywhere. Customers should not need a different privacy architecture for every data format.
03 · SCALE & ECONOMICS
THE CHALLENGE
What throughput could we credibly promise?▸ ONE CUSTOMER: 10 MB JSON FILE → 2 MIN
▸ ANOTHER: 50 × 1 MB FILES → 2 MIN
▸ ANOTHER: 50 × 1 MB FILES → 2 MIN
PRODUCT CALL
We made capacity part of the product contract.
WHY IT MATTERED
We benchmarked file size, chunking, concurrency, throughput, and GPU utilization, then converted the results into deployment sizing, customer limits, and unit economics.
04 · MODEL QUALITY
THE CHALLENGE
What evidence is enough to change a production model?
PRODUCT CALL
Model upgrades had to pass evidence gates.
WHY IT MATTERED
Annotated datasets, regression reports, and customer test cases decided readiness. A better average score was not enough if key entities or customer workflows regressed.
EXHIBIT B — BERKELEY LAB · RESEARCHER
JAN 2023 — FEB 2024
Made quantum experiments reusable instead of disposable.
THE PROBLEM
Superconducting qubits are highly sensitive to noise. QubiC generated extensive calibration and experiment data, but scientists lacked a shared way to version it, trace each result to the calibration that produced it, and collaborate across experiments.
BUILTQubiCSV, an open-source platform for storing and visualizing calibration and experiment data.
NOVELTYIntroduced Git-like data versioning so multiple scientists could contribute to the same calibration history, compare experiments, and reproduce results.
OUTCOMEHelped scientists identify better calibration settings and understand how each change affected qubit performance.
PUBLISHEDFirst-author paper in Scientific Reports, a Nature Portfolio journal, and my master's thesis.
QUANTUM NANOELECTRONICS LAB · UC BERKELEY
DILUTION REFRIGERATOR + QUBIT CONTROL SYSTEMS
DILUTION REFRIGERATOR + QUBIT CONTROL SYSTEMS
EXHIBIT C — RAKUTEN · SOFTWARE ENGINEER
NOV 2020 — AUG 2022
Built and shipped automation for global e-commerce.
SYNTHETIC DATA · 10,000+ HOURS SAVED
Built production-like test-data tooling from database structures.
CONVERSATIONAL AI · PRE-CHATGPT, 2021
Built consumer chatbot workflows using Google Dialogflow.
WORKING CONTEXT
Collaborated across India and Japan, worked cross-functionally, and mentored interns.
FIG. 03 — THE AGENT FLEET
I build agents because talking about agents is not enough.
Operating them teaches me what they are bad at, and what an agent actually needs:
CONTEXTTOOLSPERMISSIONSMEMORYSUCCESS CRITERIAREVIEW POINTSFAILURE RECOVERYA CLEAR POINT WHERE A HUMAN TAKES OVER
A PROMPT IS NOT A PRODUCT.
FLEET LOG — REPLAY OF A REAL DAY
AGENTS RUNNING
THE OPERATOR ALSO CODES —
600+GITHUB CONTRIBUTIONS ↗
200+LEETCODE SOLVED
DailyCLAUDE CODE · CURSOR · MCP
AGENT 01 — FOUNDERBUDDY
LIVE
CREW: 11 AGENTSTRIGGER: ON-DEMANDSTACK: CLAUDEHUMAN CHECK: FINAL CALL IS MINE
Your AI co-founder. An 11-agent crew.
Company research, idea pressure-testing, customer and investor simulations, pitch development, outreach prep. One Buddy rallies the right specialists. Built in two days with Claude Code, from roughly 70 minutes of voice notes. Apparently talking to myself is now a software development methodology.
TRY IT — FOUNDER-BUDDY.ONLINE ↗
AGENT 02 — RESEARCH AGENT
LIVE
SCHEDULE: DAILY 06:00OUTPUT: WHATSAPPHANDS OFF TO → AGENT 03FAILURE MODE: RELEVANT ≠ IMPORTANT
Reads arXiv while I sleep.
Five agents scan five arXiv categories every morning, rank what matters, and deliver the top 3 papers to my WhatsApp with why they matter. Approve one there, and it hands off to the Social Agent to publish.
TRY IT — RESEARCH.DAGENT.SHOP ↗
AGENT 03 — SOCIAL AGENTS
LIVE
TRIGGER: NEWS + HANDOFF ← AGENT 02CHANNELS: X · LINKEDIN · MCPFAILURE MODE: OFF-VOICE DRAFTS
Fully autonomous posting for X & LinkedIn.
Collects news, drafts in my voice, sends drafts to my WhatsApp, and schedules approved content. Also very capable of writing sentences I would never say. Human approval remains a good product feature.
TRY IT — TWITTER.DAGENT.SHOP ↗
AGENT 04 — THE AGENTIC WORK LEDGER
42s FILM
SOURCE: ~/.claude/projectsSPAN: 3 SESSIONS · 29 PROMPTSMEASURED: EVERY TIMESTAMP
Thought to reality, measured.
Three sessions of my own Claude Code logs, reconstructed end to end: 29 prompts, every tool call, every timestamp. One real loop, then all of them.
12,204:1 CACHE READS PER FRESH TOKEN
10 TOOL CALLS PER PROMPT
3 MIN WORK · 23 MIN TRUST
READ IT — THE THREAD ON X ↗
FIG. 04 — FOUNDER JOURNEY
I have started a few things.
Some of them worked. Some taught me more because they did not.
VENTURE 02 — JOBAUTO.AI · FOUNDER · MAR 2025 — MAR 2026
Browser agents that turned one click into 500 applications.
THE PROBLEM
Applying to 20 jobs properly takes 4 to 5 hours: repeated logins, document uploads, and the same questions in every system. JobAuto reduced all of it to one button.
THE QUESTION →Why should a candidate fill out the same information 20 times?
WHAT WE BUILT
The agents read unfamiliar HTML, matched every field to the user's profile, and handled login, sessions, uploads, and multi-step forms across Greenhouse, Workday, and Ashby. Fully autonomous. No review required.
LIVE DEMO — THE AGENT AT WORK
THE SIGNAL →
2,000+
USERS ON THE PLATFORM
1,000+
APPLICATIONS PER NIGHT
100K+
SOCIAL VIEWS, CREATOR PARTNERSHIPS
$5K MRR
AT PEAK
Team of 5
HIRED AND LED
1st place
AMONG 20+ FOUNDERS
WHAT BROKE →
We optimized for completed applications. Users measured success in interviews and offers.
WHAT STAYED →Measure the outcome, not the proxy. Design recovery and unit economics from day one.
LAUNCHPAD BY THEAGENTIC · PALO ALTO
VENTURE 01 — PAPERTALK.IO · CO-FOUNDER · APR 2023 — JAN 2024
5,000+ USERS
PaperTalk.io84 tabs open.
One paper deadline approaching.
One paper deadline approaching.
THE PROBLEM
While writing my Berkeley Lab paper, every useful source led to five more. Finding papers was easy. Deciding what mattered, understanding it, and citing it took hours.
THE QUESTION →Could AI help someone find the right paper, question it, and apply the research?
WHAT WE BUILTRAG-BASED RESEARCH AGENT
▸ Search — find relevant work across 100M+ research papers
▸ Ask the paper — grounded answers with supporting citations
▸ Apply the research — what could be built from the findings
LIVE DEMO · RAG RESEARCH AGENT AT WORK
WHAT BROKE →
The problem was real. Our differentiation was temporary.
Thousands tried, few paid. Native file uploads and grounded Q&A became built-in model capabilities.
WHAT STAYED →
Build for a durable workflow, not a temporary model gap.
FIG. 05 — PUBLISHED RESEARCH
Published research on the systems behind quantum computing.
FULL RECORD — GOOGLE SCHOLAR ↗
FIRST AUTHORPAPER 01 — SCIENTIFIC REPORTS · NATURE PORTFOLIO
natureSCIENTIFIC
REPORTS
REPORTS
Versioned calibration data so quantum experiments could be reproduced.
QubiCSV: open-source data storage & visualization for collaborative qubit control
DEVANSHU BRAHMBHATT · YILUN XU · NEEL VORA · LARRY CHEN · NEELAY FRUITWALA · GANG HUANG · QING JI · PHUC "VP" NGUYEN
W/ RESEARCH & STAFF SCIENTISTS @ BERKELEY LAB · PROF. PHUC "VP" NGUYEN (UT ARLINGTON)
THE PROBLEM
Quantum experiments generated extensive calibration and experiment data, but scientists lacked a shared way to version it, connect results to the calibration used, and collaborate across experiments.
MY CONTRIBUTION
Designed and built QubiCSV, an open-source platform for versioning calibration data and visualizing experiment outcomes.
WHY IT MATTERED
Scientists could reproduce experiments, compare calibration settings, and trace changes in qubit performance back to the exact calibration used.
FIG — QUBICSV: CALIBRATION DATA VERSIONING + VISUALIZATION FOR THE QUBIC STACK
THE SIGNAL →
5,611
ARTICLE ACCESSES
Top 8%i
ONLINE ATTENTION, ALL JOURNALS
Top 10%i
ONLINE ATTENTION WITHIN SCIENTIFIC REPORTS
3 outlets
NEWS IN JAPANESE, SPANISH, GERMAN
CO-AUTHORPAPER 02 — ACM MOBISYS 2025 · SIGMOBILE
acmASSOCIATION FOR
COMPUTING MACHINERY
COMPUTING MACHINERY
Mapped what fault-tolerant quantum systems demand from control, readout, and compute.
Computing systems for superconducting qubits: challenges & opportunities
W/ UT ARLINGTON RESEARCHERS · BERKELEY LAB SCIENTISTS · PROF. PHUC "VP" NGUYEN (UT ARLINGTON)
A systems-level analysis of the control, readout, leakage suppression, and computing infrastructure required to scale superconducting quantum systems.
800
DOWNLOADS — ACM DL
Open-source QubiC
CONTROL PLATFORM
FIG — RFSOC-BASED TRANSMON QUBIT CONTROL & READOUT SYSTEM
THE DEGREES BEHIND THE WORK
M.S. Computer Science
THE UNIVERSITY OF TEXAS AT ARLINGTON · 2022–24
RESEARCH ASSISTANTSHIP (RA) SCHOLARSHIP · SPRING 2023
Earned in-state tuition, reducing total tuition by approximately 50%.
RESEARCH
Master's thesis research at Berkeley Lab.
RANGE
Completed internships across AI product management, Berkeley Lab research, and software engineering.
B.Tech Information Technology
DHARMSINH DESAI UNIVERSITY · 2017–21
LEADERSHIP
Led a 30-member GDSC team and organized 20+ developer events.
FOUNDER
Co-founded Slyro while completing the degree.
ENGINEERING
Joined Rakuten during my final year and shipped software for global e-commerce.
FIG. 06 — HOW I DECIDE
How I make decisions when the answer is not obvious.
HOW I WORK
FIG. 06.1 — SIX SIGNALS → ONE DECISION
Everyone sees a different part of reality.
CUSTOMERSknow the pain.
SOLUTIONS ENGINEERSknow what blocks the deal.
ENGINEERSknow where the system will break.
RESEARCHERSknow what the model can and cannot do.
SALESknows what buyers will pay for.
LEADERSHIPknows what the company can support.
UNIT PM-01
THE PM
LISTENS. WEIGHS THE TRADEOFFS.
↓
ONE CLEAR DECISION
The PM has to listen to all of it, and still make a clear decision. That is the job.
FIG. 07 — WRITING IN PUBLIC
Writing makes the idea clear. Publishing pressure-tests it.
100K+
TOTAL IMPRESSIONS — ALL TIME
10K+
FOLLOWERS
Weekly
PUBLISHING CADENCE
SELECTED WRITING
MOST READ — LINKEDIN
Context graphs are a trillion-dollar opportunity. Labeling decisions is the trillion-dollar question.
10,000+
IMPRESSIONS
28
COMMENTS
POST — NEWLINKEDIN
Agent memory is quietly becoming the most important layer in enterprise AI.
1,000+ IMPRESSIONSMEMORY PORTABILITY UNDER GOVERNANCE
GUIDE — FEATURED BY RAGASYC W24
Enhancing LLM accuracy with RAGAS: architecture, metrics, and how to improve RAG.
FEATURED IN RAGAS' OFFICIAL DOCSGITBOOK
FIG. 08 — ASK MY AGENT
Ask what a resume cannot explain.
IT EXPLAINSwhy a decision was made, what changed, and what happened next.
GROUNDED INthe projects, papers, decisions, and postmortems across this portfolio.
IT WILL NOTinvent details, exaggerate outcomes, or answer beyond the evidence.
SYSTEM BOOT
$ ./agent.devanshu --boot
[ok] context loaded: product · engineering · research · founding
[ok] guardrails: no invention · no spin · grounded in evidence
[run] status: RUNNING — ask it anything
d
agent.devanshu
devanshu.agent ← Hi. I can explain the decisions, tradeoffs, and lessons behind Devanshu's work. What would you like to explore?
{{ m.pre }} {{ m.text }}
devanshu.agent is working…
$
SEND ↵
START HERE
▸{{ chip0.q }}
▸{{ c.q }}
FIG. 09 · REFERENCES & COMMUNITY
FORMER MANAGER · RAKUTEN
“You joined as an intern but contributed as a boundary-less professional. Given another opportunity, I will be happy to work with you again.”
KRISHNADAS C K · GROUP TECHNICAL MANAGER, RAKUTEN · DIRECT MANAGER
VIEW LINKEDIN RECOMMENDATION ↗
FIG. 10 — NEXT STEPS
Let's build AI systems people can trust with important work.
I'm interested in consequential problems across AI systems, agents, privacy, and infrastructure.
I USUALLY REPLY QUICKLY. ESPECIALLY WHEN THE PROBLEM MATTERS.
Book a call
devanshu.vguj@gmail.com
Resume ↗
DOC DB-2026 · REVIEWED
TECHNICAL AI PM · BUILDER · RESEARCHER · FOUNDER
LINKEDIN ↗
X / TWITTER ↗
GITHUB ↗
GOOGLE SCHOLAR ↗
MEDIUM ↗
BACK TO TOP ↑
SAN FRANCISCO · ENGLISH, GUJARATI, HINDI · BUILT MOSTLY THROUGH VOICE NOTES
DOC NO. DB-2026