viditparashar.me/faqFAQ

What is actually hard about building with LLMs?

Almost none of it is the model. It is the structure around it: knowing when the system should decline to answer, making a decision explainable after the fact, and choosing what happens at the edge of its competence.

The specific thing I have learned the hard way is that aggregate accuracy hides the failures that matter. A model is usually well calibrated on common cases and overconfident on rare ones — and rare cases are exactly where a clinician needed the help.

The fix is unglamorous: per-category thresholds, an explicit path to a human, and accepting a higher abstain rate rather than a better-looking headline number.

More: Clinical Agent Suite

Related results

  1. viditparashar.meviditparashar.me › about › intro

    Vidit Parashar — full-stack engineer, applied LLM and voice systems

    … AI companion and company-wide automation at Ainoviq, Blackngreen's AI division. The consistent lesson across all of it is that the model is rarely the hard part. What…

  2. viditparashar.meviditparashar.me › experience › ainoviq

    Full-stack and AI Engineer at Ainoviq (Blackngreen) (Mar 2026 – present)

    Full-timeMar 2026 – present

    1 Mar 2026 — … human-in-the-loop LinkedIn and email platform) and LinkedIn activity — these have produced real conversions. Run internal AI enablement: teaching engineers and…

  3. viditparashar.meviditparashar.me › projects › conversational-agents

    Conversational AI Platform — the layer other teams build voice and chat…

    Case studyLive2026

    1 Jul 2026 — … The people who tune an agent are not always engineers, so prompts, knowledge and voice had to be configuration rather than code. Once spoken, it is said — there is no…

    • Python
    • TypeScript
    • Next.js
    • Node.js
    • LangChain
    • LLMs
    • +4 more

Search this site for “What is actually hard about building with LLMs?”

Vidit Parashar

Full-stack and AI engineer

I build AI features for domains where a confident wrong answer costs something — mostly clinical software and voice agents.

Gurugram, India · IST (UTC+5:30) · viditsharma818@gmail.com · @viditparashar96 · in/vidit-parashar-056707225 · viditparashar.framer.website

Summary

I have spent the last two years putting large language models into production rather than into demos. Most of that has been healthcare — an ambient scribe that turns a consultation into an EHR-ready note, and a set of clinical agents deployed into university health systems — where the interesting problem is never the model, it is what the system does when it is unsure.

Experience

Selected projects

Technologies

AI and LLM systems
LLMs, LangChain, LLM orchestration, NLP pipelines, Model fine-tuning, Speech to text, Full-duplex voice, Agent orchestration, Workflow automation, Screen understanding
Languages
TypeScript, JavaScript, Python, SQL, HTML, CSS, Java
Frontend
React, Next.js, Tailwind CSS, Redux, Redux Toolkit, Zustand, ShadCN, Material UI, NextUI, Styled Components, Bootstrap, GSAP, ApexCharts, Recharts, PWA
Backend and APIs
Node.js, Microservices, Socket.io, Azure Functions, Prisma, Mongoose
Databases
PostgreSQL, MongoDB, NoSQL
Infrastructure and tooling
Microsoft Azure, AWS, Docker, CI/CD, Git, Vercel, Firebase, Appwrite, Vite, Webpack
Mobile and desktop
Swift, SwiftUI, iOS (native), ElectronJS, React Native, Ionic

Education

Availability

open to interesting problems.