SJShubham Joshi Contact
Portfolio / 2026

Applied AI + forward deployed engineer

Agents, trained models, data pipelines, and products built from first question to deployed system.

View selected work ↓ See Tacet live ↗
480+Tacet tests, engine to tenancy01 ↓ 73.1%ModernBERT validation accuracy04 ↓ 138,263rows in the ACHP master dataset03 ↓ 4AI providers in one creative system05 ↓

01 / Working range

Agent systemsApplied MLData engineeringProduct engineeringTechnical communication

Choose the evidence

One builder.
Three hiring lenses.

Applied AI engineeringSystems that plan, predict, and create.

Tacet + AgentDock + OpsPilot

Agents, models, products → Forward deployedAmbiguity converted into a client decision.

ACHP × CMS + Tacet

Discovery through delivery → Data scienceBetter decisions before bigger models.

OpsPilot + ACHP × CMS

Data, statistics, ML →

Resume index

Choose the role.
Get the relevant evidence.

The work stays the same. Each version brings the most relevant proof forward for a different hiring conversation.

01Applied AI Engineer

Tacet, AgentDock, and OpsPilot lead.

Open PDF ↗ 02Forward Deployed AI Engineer

Systems engineering, client delivery, and adoption lead.

Open PDF ↗ 03Data Scientist / Applied ML

ACHP, OpsPilot, and rigorous evaluation lead.

Open PDF ↗

02 / Selected work

Five kinds of evidence.
One operating style.

Build deeply. Validate the part that can fail. Make the result legible to the person who has to use it.

01Founder / expert judgment for AIActive / 2026

Tacet

Learns the rules an expert never wrote down, tests them on cases the expert never showed, and hands them to AI agents and new hires.

In plain English

It watches a senior operator work, guesses each decision first, and asks one short question only when its guess was wrong. The answer becomes a rule in their words that an agent can follow and a new hire can learn from.

Live early access480+ automated testsOne question: 96% → 100% held back

Founded and built it solo, from the HackNation × ElevenLabs prototype to a multi-tenant product: a plain-Python learning engine that weighs the value of a question against the cost of interrupting, sealed tests whose predictions are hashed before any human label, a voice interviewer and tutor, MCP and API exports, a Chrome extension, and a CSV import that learns a bookkeeper's rules from last month's ledger.

Live product ↗ Try the demo ↗
Fig. 01 / One rule learned
Prediction missedExpert booked it as capex
At a pause"Why capex?"Asked once
Work Map ruleNet over 3,600 → capex
Predict firstExpert's wordsSealed test
Tacet product film: Mira guesses cost centre 4711, the expert books 0400, one question becomes a rule, it passes a sealed test on eleven fresh invoices, and it stops a new hire's wrong save
Product film / 30 seconds

One rule, start to finish: the guess it got wrong, the one question it asked, the sealed test it passed, and the wrong save it stopped on Monday, in the expert's words.

What I owned

Founder, solo · Customer research, product, engine, backend, console, extension, voice agents, deployment, and test suite.

02Agent infrastructurePrototype / 2026

AgentDock

A control plane for AI agents that can use real tools while keeping permissions, budgets, and approval rules outside the prompt.

In plain English

Teams can let an AI draft, search, or act while keeping consequential steps behind enforceable rules and human approval.

225 commits61 test filesPostgres durable execution

Built the full prototype: model backed flow planning, default deny policy resolution, governed MCP execution, approval workflows, encrypted credentials, durable jobs, crash recovery, idempotent external actions, and live run events.

Live app ↗ Source ↗ YC demo ↗
Fig. 02 / A governed action
Agent intentSend report
Policy gateApproval requiredPaused safely
External toolGmail / send
Deny by defaultAuditedBudget bound
AgentDock workspace showing governed runs, approval inbox, tool usage, costs, completed work, and a halted error
Actual product / Authenticated workspace

Real governed runs expose steps, tools, cost, status, failures, and approval state. This is the operational surface behind the control plane concept.

What I owned

Solo project · Product direction, system architecture, frontend, backend, policy engine, integrations, and test suite.

03Forward deployed analyticsClient / Complete

ACHP × CMS

Turned fragmented Medicare Advantage filings into decision grade evidence for a national healthcare organization.

In plain English

I made incompatible government datasets agree, then showed ACHP where regional plans were gaining ground and where strategy needed to change.

Top three team138,263 × 208 master493 / 875 counties

Led most of the analysis within a team: built the reproducible CMS pipeline, reconciled changing schemas and geographic grain, validated enrollment logic, and translated market, quality, benefit, and plan exit signals into client decisions.

CMS data ↗ Benefits data ↗
Fig. 03 / From filing to decision
  1. 01
    IngestCMS + member files
  2. 02
    ReconcileYears, plans, counties
  3. 03
    TestCost shift evidence
  4. 04
    DecidePolicy priorities
Final consulting presentation map showing counties where ACHP plans had a competitive cost advantage
Actual result / Final consulting presentation

ACHP held the stronger cost position in 493 of 875 scored counties. The project was selected among the top three teams schoolwide.

What I owned

Analysis lead on a team engagement · Data pipeline, quality assurance, modeling, segmentation, and most analytical work. Teammates contributed research and presentation development.

04Applied machine learningModel deployed

OpsPilot

A support routing copilot where redesigning the labels mattered more than another round of model tuning.

In plain English

It reads a support request, recommends the right operating queue, and sends uncertain or sensitive cases back to a person.

52.3% → 73.1%ModernBERT7 operational queues

Trained and deployed a ModernBERT classifier, diagnosed overlapping support categories, redesigned a taxonomy of ten labels into seven operational queues, and added human review based on confidence.

Source ↗ Model ↗ Dataset ↗
Fig. 04 / The intervention
01Error analysisOverlapping queues
↓
02Task redesign10 labels → 7 operational families
↓
03Deployment ruleRoute or return to a human
Deployment behavior / Threshold 0.80
62.3%tickets routed automatically85.5%accuracy on that subset

The model abstains when confidence is insufficient; automation coverage is traded for reliability instead of forcing every prediction.

What I owned

Solo project · Dataset pipeline, taxonomy redesign, ModernBERT training, evaluation, model publication, and product integration.

05Generative productPersonal project / 2026

Vibe Curator

An AI native creative environment where generated media stays editable instead of ending as a flattened file.

In plain English

You describe a visual world, then keep shaping its image, motion, atmosphere, and music instead of downloading one finished output.

4 AI providersWebGL compositorEditable scene graph

Designed and built the product because I wanted it: generated images, music, video, and guarded shaders enter the same persistent editor with source awareness and a full screen player.

Live app ↗Watch demo ↗Source ↗
Animated Vibe Curator Beta 2 demo: create an editable moonlit scene, add atmosphere and music, then use it in Chrome or as a local desktop wallpaper Beta 2 / Live now
What I owned

Solo project · Product concept, interaction design, generative provider integrations, scene model, shader safeguards, and WebGL renderer.

03 / Additional systems

Built to answer a specific question.

Clinical AI

RxCheck

Deterministic drug interaction detection. The model explains; it never decides.

View source ↗ Voice agents

Negotiator

An autonomous voice negotiation system with explicit deal policies and constraints.

View source ↗ Consumer AI

FridgeChef

Fridge and receipt scans become structured inventory and streamed recipes.

View source ↗ Built in 14 hours

Nimbus

A burnout companion that uses tools, built during the Pava product hackathon.

View source ↗

04 / Experience

Engineering is only useful when someone can act on it.

2026 to now

Founder, Tacet

Building, solo, a product that turns an expert's judgment into tested rules for AI agents and new hires. Live in early access.

2025 to 2026

MS, Business Analytics & AI

Johns Hopkins University · Dean's Scholarship

2025 to 2026

Co founder, BAAIO

Helped grow the applied AI organization at Johns Hopkins to 100+ registered members; led technical workshops and programming.

2021 to 2024

Instructor & facilitator

Designed communication programs for adult professionals across varied backgrounds, maintaining 95%+ retention.

AcademicDean's Scholarship

Johns Hopkins University

CertificationAzure AI Fundamentals

Microsoft AI 900

Community100+ registered members

BAAIO, co founded at Johns Hopkins

Client signalTop three team

Schoolwide consulting selection; analysis presented to ACHP for strategic use

05 / About

I am most useful where product, models, data, and people collide.

I recently completed my master's in Business Analytics & AI at Johns Hopkins. My work moves between deep technical implementation and the person who needs the result. That might be an operator approving an agent action, a healthcare client defending a policy position, or a creator shaping generated media.

I have lived and worked across Japan, Australia, and the United States, and speak English, Japanese, Nepali, and Hindi at native level.

Best fit
Applied AI · Forward deployed · Founding engineer · Data science
Education
Johns Hopkins MS, completed 2026
Current focus
Tacet, my company: tested rules from expert judgment

06 / Elsewhere

Code, models,
writing, and demos.

GitHubSource and build history↗ Hugging FaceTrained model artifacts↗ LinkedInTechnical writing and updates↗ YouTubeAgentDock YC walkthrough↗

07 / Contact

Have a difficult problem
and real ownership?

shubhamjoshipro.mail@gmail.com ↗
GitHub ↗ LinkedIn ↗ Hugging Face ↗ YouTube ↗
Open to AI and data roles with meaningful ownership© 2026 Shubham Joshi