Based in Bengaluru. Available to join immediately.

AI Engineer

Nidhi Yashwanth

I build AI products that have to work with messy data, external tools, unpredictable model output, and real users.

My work spans agent workflows, retrieval, model evaluation, backend services, and the product around them. I like the awkward parts where those systems meet.

NYBengaluru, India

AI development

Agents, retrieval, and model evaluation

LangGraph and MCP workflows, RAG, structured output, and self-hosted models.

Backend and data

APIs, databases, and integrations

Python and TypeScript services, ClickHouse, PostgreSQL, OAuth, billing, and bulk exports.

Product engineering

Frontend and realtime workflows

React and Next.js interfaces, WebSockets, product state, contact tables, and user flows.

Production delivery

Deployments, bugs, and tracing

Docker, GPU services, AWS, Jenkins, monitoring, and debugging across service boundaries.

Selected work

What I built, and what changed.

01 · Agent and tool systems

A lead generation platform that researched companies and enriched 30K records in one request.

I worked on the research agent and the product around it, including enrichment providers, contact tables, exports, and CRM handoffs.

6

TypeScript MCP servers

~57

tool operations

30K

records in one bulk request

  • Built the services for enrichment and verification, including a bulk path that streamed multi-million-row CSV exports to S3 with constant memory.
  • Connected the LangGraph research flow to a runtime tool registry, then gave it natural-language access to revenue data split across four ClickHouse tables.
TypeScript
MCP
LangGraph
ClickHouse
S3

02 · Retrieval

2,000 support conversations became a typed knowledge base.

The support history held answers that the product could not reliably retrieve. I built the path that made them searchable and inspectable.

2,000+

conversations processed

200 → 1,250

useful chunks

17

product areas covered

  • Built the extraction and ingestion path with typed JSON output, Zod validation, and repair for malformed model responses.
  • Stored the result in Postgres with pgvector and expanded the useful corpus more than 6x.
Python
PostgreSQL
pgvector
Zod
RAG

03 · Model evaluation

A 35B model got an eval loop before production.

I needed to know where a local model was right, where it failed, and whether the next prompt actually improved anything.

188 tok/s

quantized 35B inference

101

companies in the audit set

45% → 73%

severity accuracy

  • Ran a quantized MoE model through llama.cpp and CUDA, then built a regex-first verification flow around it.
  • Against a hand-reviewed set, phone extraction moved from 15% to 73% and critical-data detection reached 97%.
llama.cpp
CUDA
Python
Evaluation
GGUF

04 · Delivery and realtime systems

90 minutes became 15. Four minutes became 50 seconds.

I worked on release paths in two earlier roles: GPU services at one, Jenkins deployments at another.

6x

faster GPU deployments

~5x

faster Jenkins deployments

500+

weekly new users on realtime flows

  • Containerized GPU inference and cut one AI service deployment from roughly 90 minutes to 15.
  • Moved Jenkins builds to pre-built Docker images in ECR, cutting another release path from 3-4 minutes to about 50 seconds.
  • Reworked chat and profiling state with Socket.IO, WebSockets, and Redux for a product adding 500+ users each week.
Docker
Jenkins
AWS ECR
WebRTC
Socket.IO

Independent work

Things I built because I wanted them to exist.

A fragrance store, a commercial gallery concept, a profit model, browser experiments, a community, and tools I wanted enough to build myself.

SEREIN

July 2026

SEREIN

A fictional fragrance store built as a complete client project: campaign-led homepage, product catalog, detailed fragrance pages, scent finder, persistent bag, and a clear order path.

  • Built nine responsive routes around one store, with original campaign art, a three-question scent finder, local bag state, structured product data, canonical URLs, and sitemap coverage
  • Built a native WebGL2 ray-marched 3D bottle with direct manipulation, momentum, responsive glass and liquid materials, scroll-controlled camera movement, and reduced-motion behavior
Next.js
TypeScript
WebGL2
GLSL
Ray Marching
Modern CSS
Enter the store
LOAM Editions

July 2026

LOAM Editions

A fictional collectible design gallery built like a real client project: a cinematic front page, priced collection, product stories, viewing appointments, and a clear route from interest to inquiry.

  • Built nine responsive routes around one retail story, with original product art, structured product data, canonical URLs, and a complete sitemap
  • Used native WebGL2, GLSL, scroll-driven CSS, and cross-document view transitions as progressive enhancement around familiar buying and inquiry paths
Next.js
TypeScript
WebGL2
GLSL
CSS View Transitions
Enter the gallery
MARGIN

July 2026

MARGIN

An interactive commerce profit model for the question most dashboards dodge: what does the store keep? Change conversion, order value, media spend, returns, and fulfillment costs and the full model responds.

  • Tied a native WebGL2 fragment shader to the live operating margin instead of using motion as decoration
  • Added break-even ROAS, cost of delay, keyboard controls, reduced motion, metadata, and social previews to the same static route
Next.js
TypeScript
WebGL2
GLSL
CSS scroll timelines
Run the model
Latent Space Observatory

July 2026

Latent Space Observatory

An interactive field study of how context becomes attention and attention becomes language. I wrote the WebGPU/WGSL renderer and its WebGL2/GLSL fallback directly, without putting a 3D engine between the idea and the GPU.

  • Connected pointer movement, scroll progress, and sampling temperature to the live shader
  • Built the token inspector, attention matrix, and reduced-motion path around the same field
React
TypeScript
WebGPU
WGSL
WebGL2
GLSL
Open experienceSourceRead the concept

June 2023 - Present

CodeIndia.tech

I co-founded a learning community that grew to 15K+ members and about 1K weekly sessions. I built the product and kept the Node/MongoDB backend in shape as it grew.

  • Handled the software and the day-to-day work of growing the community
  • Kept common paths under a second with query and caching work
Next.js
Node.js
MongoDB
Azure
Visit

sitesdevshouldknow.com

2025

sitesdevshouldknow.com

I kept a list of useful developer sites. It got out of hand, so I turned 1,000+ links across 27 stacks into a searchable Postgres-backed product.

  • Built the relational data model and server-rendered search surface
  • Added authentication and a simple way to keep the directory organized
Next.js
TypeScript
PostgreSQL
Prisma
Visit

January - February 2024

LLMSpace

Built a searchable index of 500+ AI tools during the 2024 tool rush. Most of the work was ingestion, organization, and making a crowded category easier to browse.

  • Built the ingestion APIs and search model
  • Built and published the first version over January and February 2024
Next.js
Node.js
MongoDB
Tailwind CSS
Visit

How I work

Follow the whole path.

I do my best work when I can understand the product, trace the data, and stay with the feature through release.

I move between React screens, LangGraph services, ClickHouse queries, GPU containers, OAuth, billing, and release pipelines when the work calls for it.

Trace the system

I follow a request through the UI, model, tools, storage, and logs until I can point at the failing boundary.

Product flow · Data contracts · Failure paths

Measure model behavior

I build small test sets before tuning prompts or swapping models. They stay small enough to review by hand and rerun after each change.

Eval sets · Grounded outputs · Regression checks

See the feature through

I work across the screen, API, data, and release path. That means handling edge cases, fixing what breaks in testing, and staying with the feature until people can use it.

Frontend and backend · Bug fixing · Release

Working set

Tools I have used in shipped work.

AI systems

  • LangGraph
  • MCP
  • RAG
  • Model evaluation
  • llama.cpp
  • pgvector

Backend and data

  • Python
  • TypeScript
  • NestJS
  • FastAPI
  • ClickHouse
  • PostgreSQL

Product and delivery

  • Next.js
  • React
  • WebGL
  • Realtime systems
  • OAuth
  • Docker
  • AWS / Jenkins

Experience

A short work history.

Worked across the lead generation product, its research agent, retrieval pipeline, and customer-facing data workflows.

Built six TypeScript MCP servers with about 57 tools and a 30K-record bulk path that streamed multi-million-row CSV exports to S3 with constant memory.

Turned 2,000+ support conversations into about 1,250 typed pgvector chunks across 17 product areas, with validation and repair for malformed model output.

Added natural-language analytics over four ClickHouse tables, model evaluation on self-hosted hardware, OAuth integrations, Stripe billing, exports, and monitoring.

Authored 118 commits across refs and touched 157 files.

Built an agentic AI interview platform in Python and NestJS, with A2A coordination, MCP tool use, and reasoning-model candidate ranking.

Developed v2 of the AI proctoring system as a 40+ endpoint backend for realtime WebRTC audio and video, fraud detection, and monitoring.

Cut backend fetch time by 25% with bounded pagination using limits and offsets.

Containerized GPU inference and cut deployment time from about 90 minutes to 15 with the NVIDIA Container Toolkit.

Built onboarding for profiling, education, and work history in a mobile-first Next.js product.

Implemented realtime chat and role flows with WebSockets, Redux, optimistic updates, infinite scrolling, and persisted state.

Reworked data flow for 500+ weekly new users across the Next.js app, Strapi/PostgreSQL backend, and Flask/Gemini service.

Moved Jenkins releases to pre-built Docker images in ECR, cutting deploys from 3-4 minutes to about 50 seconds.

Background

Education and credentials.

Buildspace

s4, s5 · 2023-2024

Government Engineering College, Gangavathi

BE, Computer Science · VTU · 7.5/10 GPA · 2019-2023

Azure AI Engineer AssociateMicrosoftAzure Data Scientist AssociateMicrosoft

420 leetcode problems; 1,625 peak contest rating.

Writing

Notes from building with models.

I write when a benchmark, implementation detail, or security boundary is worth keeping around.

Running OpenRouter coding models through Claude Code

How I routed Claude Code through OpenRouter, selected coding models, and made the setup practical on Windows.

July 26, 2026 (1d ago)

Learning WebGL through a fragment shader

What I learned from treating the canvas as a programmable surface instead of a place to draw DOM-like objects.

July 25, 2026 (2d ago)

How a language model chooses its next token

Learn how tokenization, embeddings, attention, logits, and temperature produce the next token in a large language model.

July 19, 2026 (1w ago)

Review MCP servers for tool boundaries and credential risks

A practical MCP server review method based on protocol boundaries, authority, schemas, credentials, and safe defaults.

July 11, 2026 (2w ago)

Set up Stripe subscriptions in a Next.js SaaS app

A practical Stripe subscription design that uses Checkout, webhooks, the customer portal, and a small local cache.

April 19, 2026 (3mo ago)

Gemma 4 moves agent workflows onto edge devices

What Google's Gemma 4 and LiteRT-LM release means for tool-using agents on phones, laptops, browsers, and small edge devices.

April 2, 2026 (3mo ago)

Qwen3.5-9B on an RTX 4060: Ollama vs. llama.cpp

A local Qwen3.5-9B benchmark on an RTX 4060 laptop, with measured throughput from Ollama and llama.cpp.

March 31, 2026 (3mo ago)
Browse all writing

Contact

Tell me what you are trying to ship.

If my background fits what you are building, send me the context, the deadline, and the part that is currently stuck. I read every message myself.

EmailWhatsAppLinkedIn
Use WhatsApp
LinkedIn
X