// AI ENGINEER · DATA ANALYST
VATSAL DOSHI
I build LLM/RAG systems — and the analytics that prove they work. Currently shipping recommendations to 50K+ daily queries as AI Engineer at Connyct.
four films, one diagnosis. choose.
DAILY QUERIES SERVED BY A REC ENGINE I TUNE AT CONNYCT
CITATION ACCURACY ON A RAG SYSTEM DELIVERED TO ELASTIC — UP FROM 90%
PARTNER DOC SEARCH TIME — DOWN FROM 10+ MINUTES
MANUAL REPORTING TIME CUT FOR A DC NONPROFIT SERVING 1,500+ STUDENTS
The AI proposes.
Deterministic code executes.
And every number gets a defense.
I'm Vatsal — an engineer from India, via an M.S. at the University of Maryland, now working remotely from Chicago. Days go to Connyct, where I run the AI features of a short-video app for college students; nights go to building tools like Switchboard and Marquee — a movies + TV tracker I built because I log everything I watch.
Off the clock: Formula 1 race weekends, planning trips around roller coasters (the enthusiast kind of enthusiast), and a boba habit I refuse to quantify — the one number that doesn't get a dashboard.
I work the full loop: user interviews → SQL/pandas analysis → LLM systems in production → dashboards that tell leadership the truth. My projects share one rule — models are never trusted with what code can verify. Retrieval gets test suites, recommendations cite their sources, and migrations ship with rollback plans.
NOW
AI Engineer @ Connyct
Short-video social platform for college students, with a Warner Music Group content partnership. Scaled the content recommendation engine to 50K+ daily queries; 92% user satisfaction across AI features from 50+ user interviews.
AI Engineer @ Elastic — consulting engagement, UMD
Built and delivered the Partner-Docs RAG assistant below — hybrid retrieval, 50–60% latency cut, a 25-case accuracy suite, and iterative delivery with Elastic client stakeholders.
Data Analyst Intern @ Tech Turn Up
Sole data analyst for a DC nonprofit teaching 1,500+ K-12 students. Built the analytics stack from scratch — 8 Tableau dashboard pipelines, 75% less manual reporting, and cohort analysis that drove an executive restructuring.
2025
M.S. Information Systems @ University of Maryland
College Park, MD. Coursework across data analytics, database management, designing AI systems, and cloud computing & big data. Certified ScrumMaster; Google Project Management Certificate.
2024
B.E. Computer Engineering @ Gujarat Technological University
Ahmedabad, India — CGPA 3.8/4.0.
-
001
DAYS → <3 MINSwitchboard ↗
AI data-migration copilot. Legacy CRM export in, clean import out — the LLM proposes mappings and consultant-style questions; deterministic pandas executes. Days of onboarding → under 3 minutes.
-
002
URL → DEMO IN ~37sColdOpen ↗
AI demo engine for sales teams. Paste a prospect's URL → a tailored interactive demo, a 4-beat talk track, and an ROI brief — bound together by a single sales thesis so they never contradict.
-
003
TASTE, EXPLAINEDMarquee ↗
Movies + TV tracker with an AI taste engine. Recommendations cite the films of yours they reason from, match scores are honest, and every AI suggestion is verified against TMDB before it renders.
-
004CLIENT WORKNO PUBLIC ARTIFACT90% → 95%+ CITATIONS
Partner-Docs RAG FOR ELASTIC (NYSE: ESTC)
Production-grade document assistant for Elastic's global partner program. Hybrid BM25 + vector retrieval over 1,900 chunks, page-level citations, validated across 5 LLM providers with client stakeholders.
-
005
RESUME → VERDICT IN ~30sCareerMatch ↗
RAG resume-vs-JD analyzer — upload a resume, paste a job description, get a weighted score, gap analysis, and the missing ATS keywords as a PDF report. Scored against an explicit rubric, not vibes.
-
006
PLANNER → RESEARCHER→ ANALYST → WRITERLIVE AGENT FEEDQUESTION → CITED REPORT
Multi-Agent Research Assistant ↗
Four agents split a research question, search the web, run analysis code in a sandboxed subprocess, and return a cited report — with a live Streamlit feed of them working. Citations travel as state, so the Writer can't cite what the Researcher never found.
-
007
10,000+ PAGES INDEXED<5% HALLUCINATION<600ms ON CACHE HIT+35% RELEVANCE vs PURE VECTOR
CodeQuery ↗
Ask Python library docs questions in plain English — answers come only from the actual documentation, with cited pages, so deprecated parameters and invented methods never reach you. Hybrid retrieval with semantic caching on Redis.
One résumé,
compiled to target.
Same career, same numbers — weighted for the role you're actually hiring for. Pick a target and it builds. No email gate, no form.
REV 2026.07 · 1 PAGE · EVERY CLAIM SOURCED
// SELECT COMPILE TARGET
OPEN TO DATA / AI ANALYST & RAG-LLM ENGINEER ROLES
vatsal.doshi.cs@gmail.com →prefer a CLI? press [/] and type 'help'. there are secrets.