Shivam Bhardwaj
New Delhi IST

Shivam Bhardwaj

Fig. 0

AI engineer. Agents, evals, and the interfaces around them.

Role
AI Engineer
Now
Nutricheck, remote
Based
New Delhi, India
Inks

Fig. —

§01Index

I build AI systems, and the product around them.

I design and build AI systems end to end: the pipelines, the agents, the evals that keep them honest, and the product people actually use. I started in design, so both halves matter to me.

5+
years shipping production software
3
agent-evaluation tools built in 2026
2
open-source packages, PyPI & npm

Fig. —

§02Systems

An agent is a system, not a prompt.

  1. 2.1Context

    Ground it in real data.

    Retrieval over real data, with sources you can point to.

    RAG / pgvector / Chroma / FHIR

  2. 2.2Orchestrate

    Make the flow explicit.

    LangGraph state machines, not one heroic prompt.

    LangGraph / routing / multi-agent / state

  3. 2.3Act

    Tools, with a human on the consequential ones.

    Agents act; anything irreversible waits for a human.

    tool use / human-in-the-loop / Bedrock / Strands

  4. 2.4Evaluate

    Evals before demos.

    Judges score it, deterministic gates block it, baselines catch regressions.

    LLM-as-judge / regression gates / CI

  5. 2.5Ship

    Stream it, watch it, keep it cheap.

    Streamed, observed, and routed to the right model.

    SSE / streaming TTS / observability / cost routing

A replayed run

can-i-say-yes · investigator runqueued0.0s

› Can we deliver the Northwind spring campaign by Friday?

  1. parse_requestorchestrateintent: delivery promise · deadline: Fri · scope: campaign
  2. get_commitmentscontext
  3. check_capacityact
  4. check_suppliersact
  5. read_calendarcontext
  6. schedule_gateevaluate
  7. verdictship
  8. await_approvalact
  9. monitororchestrate
  10. supplier_delaycontext
  11. re_evaluateevaluate

A reconstruction of the Can I Say Yes? demo flow, abridged. Names and timings are illustrative.

Fig. —

§03Selected work

Things I built to find out how.

  1. 3.12026Python CLI & SDK · PyPI

    EvalRun

    A quality score is not a release decision.

    Local-first testing for AI agents: LLM judges, hard policy gates, regression baselines, one CI exit code.

    Python / LLM-as-judge / vLLM / Ollama / CI

    GitHubPyPI

  2. 3.22026WebGL engine · npm

    Scree

    One field of points, any shape.

    A WebGL engine that morphs one field of points between images, words and 3D. This page runs on the same idea.

    TypeScript / WebGL / GLSL / Vite

    LiveGitHubnpm

  3. 3.32026Release assurance · OpenAI Build Week

    Prism

    Evidence before an agent goes to a customer.

    Compares agent releases on a versioned suite and records who approved what.

    Python / SQLite / Agents / Policy engine

    GitHub

  4. 3.42026Agent · Agents for Humans

    Can I Say Yes?

    Before you promise a customer, let the agent check reality.

    An investigator agent answers SAFE, UNSAFE or UNKNOWN with cited evidence, then keeps watching.

    Strands Agents / AWS Bedrock / Python / Next.js

    LiveGitHub

  5. 3.52026Multi-agent studio · Gemini Live Agent Challenge

    DreamFrame

    Tell a story, watch it get drawn.

    A director agent and five specialists turn a prompt, voice or photo into a streamed graphic novel.

    FastAPI / Gemini / Live API / React

    GitHub

  6. 3.62026Orchestration · observability

    LLM Gateway Lite

    Route, execute, evaluate, recover.

    Gemini-first routing, a LangGraph pipeline and sampled LLM-judge evals, with an observability UI.

    NestJS / LangGraph / pgvector / Redis

    GitHub

Writing

Also in the drawer

Everything else on GitHub

Fig. —

§04Record

Five layers, newest on top.

Full CV (PDF)

  1. Feb 2025 — PresentRemote

    NutriCheck

    Senior Full Stack / AI Systems Engineer

    • LangGraph pipelines for lab reports, biomarkers and nutrition plans.
    • Voice-first health agent with streaming TTS; FastAPI and NestJS services.

    LangGraph / FastAPI / NestJS / Next.js / Vue / FHIR / Stripe

  2. Feb 2025 — Jun 2025Remote · Contract

    KIWIQ.ai

    Founding Frontend Engineer

    • Rebuilt an AI content platform and its workflow builder.
    • LLM pipelines turning LinkedIn data into content strategy.

    Next.js / LLM orchestration / Design systems

  3. Sep 2023 — Dec 2024New Delhi

    QPe · Appernity

    Full Stack Engineer

    • Next.js dashboards for 100K+ monthly visitors.
    • Five client products across React, Vue, Ember and Ruby.

    Next.js / Django / Ruby / AWS / GCP

  4. Aug 2022 — Aug 2023Noida

    DropTheQ

    Frontend Developer

    • Load time from 3.0s to 2.5s; 4.5/5 user satisfaction.

    React / Performance

  5. Dec 2020 — Jul 2022Remote

    FlixStock

    UX/UI Designer → Frontend Developer

    • Three UI overhauls in six months, from Figma to production.

    Figma / React / Prototyping

Fig. —

§05Tools

The workbench, roughly by reach.

Applied AI

LangGraph, LangChain, RAG pipelines, Pinecone, FAISS, Weaviate, OpenAI, Gemini, Bedrock, PyTorch, Agent evaluation

Backend

Python, FastAPI, Django, Node.js, NestJS, Go, PostgreSQL, MongoDB, Redis, GraphQL, Kafka

Frontend

TypeScript, React, Next.js, Vue, Nuxt, WebGL, Framer Motion, Zustand, React Native

Infrastructure

Docker, Kubernetes, AWS, GCP, CI/CD, WebSockets, Stripe, FHIR

Fig. —

§06Contact

Building something that has to hold up?

Open to AI engineering roles and projects.

GitHubLinkedInCV