Hi, I'm Yash Shah 👋

💻 CS + Data Science @ UC Berkeley

Software Engineer · Data Science · AI / ML

Full-stack apps, scalable backend systems, ML pipelines, and LLM-powered workflows.

class YashShah:
    def __init__(self):
        self.school   = "UC Berkeley"
        self.degrees  = ["Computer Science", "Data Science"]
        self.focus    = ["Backend Systems", "Data Infrastructure", "AI / ML"]
        self.stack    = ["Python", "TypeScript", "Java", "SQL"]
        self.building = "agent systems with deterministic guardrails"

    def philosophy(self):
        return "If a number shows up on a screen, you should be able to trace it."

• 🔭 Currently building sales-engine and relay — systems where a model does the reading and the reasoning, and deterministic engines stay in charge of what is true. Both have live demos.
• 🌱 Going deep on distributed systems, LLM evaluation, and data infrastructure.
• 💬 Happy to talk about FastAPI, Postgres, retrieval pipelines, agent evals, or SQL that refuses to be slow.
• ⚡ Fun fact: most of my side projects start because I wanted to check a number somebody else asked me to trust.

Connects product signals, enrichment and CRM state to work out which accounts deserve attention, why now, and who owns them. The interesting part is what it refuses to claim: the lead-scoring model posts a structural AUC of 0.537, not distinguishable from random, and the repo leads with that instead of the flattering leakage-contaminated 0.593. The best-performing subject line lifts replies by 16.3 points and is still recommended against, because the extra replies are angry ones. One of four integrations is verified by execution, and the other three say so in the product.

Ten Harbor tasks built from realistic inherited data-science work: a business memo, the messy warehouse behind it, the document that governs the decision, and an analysis someone already did that runs, is well documented, and answers the wrong question. Both frontier models tested reconstruct the evidence and frame the right question at or near ceiling; producing the decision-relevant number is the floor for both, and the instrumented trials break into eleven distinct failure patterns rather than one. Every arm graded by the same frozen verifiers.

An adaptive Claude agent reads a messy industrial request and decides what to investigate. Deterministic engines decide what is true: compatibility, stock, pricing and approvals never come from the model, and any claim the engines did not produce is rejected before it reaches a customer. 14 of 14 business outcomes correct in the eval suite, with zero safety violations.

Takes messy legacy ledgers through validation, reconciliation and a governed change workflow, then answers the only question that matters: is this migration safe to launch? 37 deterministic rules and 10 reconciliations turn 65 findings into the 20 decisions a person actually has to make, every fix needs an approver who did not request it, and 12 readiness gates sit on a hash-chained audit trail. Switching the AI off removes no correctness.

A mid-sized knit-fabric business runs on spreadsheets, WhatsApp and the owner's memory. TextileOps does the noticing continuously: a deterministic exception engine re-derives ten kinds of problem from current state, an impact engine attaches days at risk and revenue exposure to each one, and every exception carries its evidence down to the supplier's own message. The arithmetic is tested code; the model only reads unstructured material and drafts explanations.

One connected system for the whole lifecycle: application, screening, qualification, onboarding, invitation, staffing, work review, payment preparation and offboarding. Routine coordination is automated and consequential decisions are not — ten named approval gates require a signed-in human, and each one is recorded against the operator who made it.

A simulator and evaluation toolkit for enterprise support agents. 190 frozen scenarios across three support and billing workflows, exact offline replay, and trace-aware safety grading with no LLM judge in the loop.

Built around the problem of understanding transactions and charges across accounts and cards. Receipt OCR, explainable alerts, and verified financial analysis over synthetic data.

Internal GPU infrastructure operations dashboard with a transparent, rule-based capacity allocation planner. Every allocation decision is explainable.

Multiple users editing the same document at once, with conflict resolution handled by CRDTs rather than locks.

Students rate and review dorms, filter by college, and browse community feedback.

Predicting cricket match results from historical data, with feature engineering on team and venue history.

🤝 Let's build something

I'm open to internships, research collaborations, and interesting problems in backend, data, and applied ML.