Available for freelance work Jaipur, India · IST

Yashvardhan Thanvi.

Founder of LLMSlim, the open-source context layer for production AI agents, backed by 14 startup programs including OpenAI, Claude and Sarvam. I also run AMEYOR, my independent AI and software studio. Previously AI/ML research at DRDO.

Start a project Watch the LLMSlim demo Resume
GitHub LinkedIn YouTube Instagram founder@llmslim.app

LLMSlim scored 78.7 on NaturalQuestions at a 25 percent token budget, level with Microsoft's LongLLMLingua at 78.0 and ahead of LLMLingua-2 at 69.0, while compressing about 100 times faster.

+9.7 pts
ahead of Microsoft's LLMLingua-2, same token budget
14
startup programs backing LLMSlim
Rank 1
Rajasthan, Viksit Bharat Young Leaders Dialogue 2026
6 merged
PRs in Pydantic, NVIDIA NeMo, ModelScope, IBM, AWS Labs

LLMSlim is backed by 14 startup programs

  • OpenAI for Startups
  • Claude for Startups
  • Sarvam Startup Program
  • Zoho for Startups
  • MongoDB for Startups
  • Auth0 for Startups
  • Zendesk for Startups
  • Mixpanel for Startups
  • Sentry for Startups
  • Descope
  • Pulumi for Startups
  • Honeycomb for Builders
  • Exa
  • Neo4j Startup Program

01 / Ventures

Two companies. One builder.

LLMSlim is the product: context infrastructure that AI teams drop in front of every model call. AMEYOR is the studio: my sole proprietorship that ships LLMSlim and Krishyak, and builds for clients with the same engineering.

LLMSlim

Founder & developer · 2026 to now

Core v0.9.0 on PyPI
LLMSlim: the context layer for production AI agents Watch the demo · 1:46

The context layer for production AI agents. LLMSlim decides what reaches the model, keeps provenance on every piece, and fits it all into your token budget, locally, with no API key. An MIT open-source core on PyPI, plus a private Platform in beta with traces, replay and an auditable memory timeline.

78.7vs 69.0 for Microsoft's LLMLingua-2NaturalQuestions · 25% budget
~100× fasterlevel with LongLLMLingua at 78.00.09 s vs 9.3 s per question
62× cheaperper question on long chat historiesLongMemEval · 14× fewer tokens
3× fewer tokenson long documents, same scoreLongBench v1 · 54.9 vs 54.9
  • Product Hunt #11 of the day
  • 12 PyPI releases
  • 900+ automated tests
  • 16 GitHub stars
  • LangChain · LlamaIndex · Haystack · MCP
llmslim.app Source PyPI Benchmarks Product Hunt

AMEYOR

Founder · Sole proprietorship · Jaipur

Taking new projects

Engineering what comes next. AMEYOR is my independent technology studio and a Udyam-registered MSME. It builds intelligent software, practical AI and digital products, including its own, LLMSlim and Krishyak. Then it brings the same engineering to client work.

Written scopes before anything is built, published starting prices, reviewable milestones with demos, and one accountable engineer from brief to handover. The person who scopes your project builds it.

  1. 01Brief
  2. 02Scope & quote
  3. 03Kickoff
  4. 04Build
  5. 05Handover
  • Websites and web apps
  • AI and ML engineering
  • Custom software and SaaS
  • Agentic workflows and automation
  • UI/UX and product design
  • Consulting and maintenance
1,608automated tests across both products
2 daystypical reply time to a new brief
Projects and enquiries
hello@ameyor.in
ameyor.in Pricing Estimate a project Book a call

02 / Selected work

Shipped, benchmarked, demoed.

Products in production, research prototypes and international hackathon builds. Each one comes with a live app, a demo film or both, and numbers you can check.

Krishyak field intelligence app Product film · 1:47 Live in production
AgriTech · AMEYOR product2025 to now

Krishyak

Field intelligence for Indian farmers. Crop-photo disease checks, Sentinel-2 satellite views, local weather, mandi prices, government schemes and what-if planning, in English plus all 22 scheduled Indian languages. Built for low-end phones and patchy networks, with passkey sign-in live in production.

  • 93.71% held-out accuracy on 8,566 photos across 38 classes and 14 crops
  • 98.5% accurate on the photos it answers; it abstains when unsure instead of guessing
  • Presented to the Union Agriculture Minister and the DG, ICAR; 1st place at IETE Engineers' Day 2026
  • 700 automated tests across the FastAPI backend and Next.js frontend
  • Next.js 16
  • FastAPI
  • PostGIS
  • LiteRT
  • Sarvam
  • Sentinel-2
Open the app Demo farm Source
Pratibimb digital twin landing page Live
HealthTech · Team leadDigital Twin Challenge 2026

Pratibimb

A living digital twin of a person with Type 2 diabetes. Personalised from a FHIR R4 health record, synced to every CGM reading by a 64-member Ensemble Kalman Filter on a physiology model, and forecasting glucose four hours ahead with every alert explained.

  • 0.986 AUROC for one-hour hypoglycaemia alerts
  • 29.8 mg/dL 30-minute RMSE vs 41.9 for linear extrapolation; 93% in Parkes zone A
  • EnKF
  • LightGBM
  • TreeSHAP
  • FastAPI
  • Next.js 15
Open the app Source
Hearthbeat caregiver dashboard with an Alexa check-in Watch the demo
Alexa+ · MCP · Privacy-firstAmazon Developer Hackathon

Hearthbeat

Care that asks before it alerts. A companion for older adults who live alone that learns a home's daily rhythm from Ring motion and door sensors, never video or audio. When a day looks off, Alexa+ checks in by voice first; family is alerted only if needed, always with the reasons.

  • 14 ms p50 MCP response over Streamable HTTP; Alexa+ asks for under 500 ms
  • 12-tool MCP server, signed Ring webhooks, OAuth 2.1 and an open-sourced dayrhythm library
  • Python
  • MCP SDK
  • Alexa+
  • React
Demo Source
Polygraph: every pull request makes claims Demo
Dev tools · AgentsNVIDIA

Polygraph

Every pull request makes claims. Polygraph runs base and head side by side in isolated universes, builds a claim ledger from the PR description and flags undeclared behaviour changes, each with a reproducible witness and a committable pytest.

  • $0.03 to $0.06 per full examination on NVIDIA Nemotron
Live replay Source
Passerelle fact-lock replacing figures with placeholders Demo
Sovereign AIHack Apertus

Passerelle

Every letter, understood, without leaving the building. An air-gapped caseworker assistant on Apertus 8B that explains Swiss official letters in the client's own language. Fact-lock placeholders mean the model never sees the digits.

  • 89% of critical facts exact across 10 languages, vs 55% naive
  • 98.9% fact-extraction micro-F1
Demo Source
StudyTwin answering from a cited lecture slide Demo
Multimodal RAGIIT Mandi

StudyTwin

A tutor that shows its sources. Lectures, slides and textbooks become a cited knowledge base where every answer opens the exact page, slide or timestamp, and off-material questions are politely declined.

  • 0.97 RAGAS faithfulness; 100% of answers cite a gold location
  • LLMSlim trims prompts 30 to 38% with no loss in grounding
Demo Source
CanopyProof: is the forest behind your credit still standing? Watch the demo
ClimateTech · On-chain proofsIEEE ClimateChain 2026

CanopyProof

Satellite-verified integrity for forest carbon credits. Sentinel-2 change detection becomes signed, content-hashed records, anchored on-chain in monthly Merkle roots that anyone can re-verify from the original pixels.

  • 99.6% precision, 99.9% recall against Hansen forest loss on the Bootleg Fire
  • 1,132 signed records in 82 monthly batches
Demo Source
RuleDost: hackathon fine print, decoded with receipts Live
Developer tools · No AI at runtimeBuild With AI 2026

RuleDost

Hackathon fine print, decoded with receipts. Paste a rules page and get the real deadline on your own clock, your eligibility and a submission checklist, each backed by the exact source sentence. Runs entirely in the browser.

  • Found the deadline on 39 of 39 real Devpost pages
  • 54 unit tests, no server, .ics reminders
Open the app Demo Source

More builds

  • BharatVaaniMultilingual AI news with summaries, translation and spoken articles in 12+ Indic languages.
    DemoSource
  • GlycoTwinType 2 diabetes twin on real CGM data. Flagged 81% of 2,783 glucose spikes a median 60 minutes ahead.
    Source
  • LeashEnforceable spending mandates for AI shopping agents, settled on PayPal. Built for the PayPal AI Hackathon 2026.
    Source
  • EngageXAn interactive study of what keeps people playing, built on Steam data and machine learning.
    LiveSource
  • IntuixExoplanet detection with machine learning. Global Finalist Nominee, NASA Space Apps Challenge 2025.
    Source
  • PRAGYAANCivic intelligence platform by Team Intuix for India Innovates 2026.
    Demo

03 / Open source

Fixes that ship upstream.

I contribute to the tools the AI ecosystem runs on: 50+ pull requests across 35+ projects, from Pydantic and NVIDIA NeMo to docling, crewAI and FAISS. Merged work is below; the rest is in review.

6

pull requests merged into projects with up to 4,500 stars

52

pull requests opened to external repositories

GSSoC '25 contributor Open Source Connect India '25 LLMSlim maintainer
All merged PRs
  1. Treat logfire.tags passed through logging and loguru as real log tags
    pydantic/logfire #2550★ 4.5k
    Merged
    Oct 2026
  2. Keep context length and depth as metric dimensions in needle-in-a-haystack reports
    modelscope/evalscope #1851★ 3.5k
    Merged
    Oct 2026
  3. Name numeric dimensions in metric labels across the CLI, HTML report and web UI
    modelscope/evalscope #1856★ 3.5k
    Merged
    Oct 2026
  4. Break the models.errors import cycle that failed on first import
    NVIDIA-NeMo/DataDesigner #1039★ 2.3k
    Merged
    Oct 2026
  5. Add an opt-in raise mode so LLM failures surface instead of silently falling back
    IBM/FactReasoner #63
    Merged
    Oct 2026
  6. Remove 610 lines of unreachable ConfusionMatrixCalculator code
    awslabs/stickler #390
    Merged
    Oct 2026

In review at

  • docling ★69k
  • crewAI ★60k
  • FAISS ★41k
  • Lexical ★24k
  • Prophet ★20k
  • NVIDIA SkillSpector ★20k
  • Microsoft SkillOpt ★18k
  • NLTK ★15k
  • Pillow ★14k
  • IBM MCP Context Forge
  • BoTorch
  • Mozilla any-llm
  • NVIDIA NeMo RL
  • LibCST
  • libpng
  • MSAL Python
  • OpenTelemetry

04 / Experience

Research, product and the people around it.

  1. 2026 to now

    Founder & Developer

    LLMSlim

    Built LLMSlim end to end: the compression engine, Adaptive Context Planner and Agent Context Runtime, plus the website, Studio, docs and public benchmarks. Launched on Product Hunt and grew it into 14 startup programs.

  2. 2026 to now

    Founder

    AMEYOR · Sole proprietorship, Jaipur

    An independent studio for web, AI and custom software. Runs LLMSlim and Krishyak as in-house products and delivers client builds with written scopes and milestone demos.

  3. May to Jul 2026

    AI/ML & Software Research Intern

    Scientific Analysis Group, DRDO · Ministry of Defence, New Delhi

    Built offline, air-gapped RAG pipelines with hybrid lexical and semantic retrieval, cross-encoder reranking and local LLM inference, plus prototypes that synthesise and validate policy-compliant hardware configurations. Supported software-defined radio research on intermittent jamming and out-of-band spectral behaviour.

  4. Mar to Jun 2025

    AI/ML Engineering Intern

    Olinp Technologies · Remote

    Refactored preprocessing and feature-engineering pipelines across 3+ production modules and shipped optimised ML services behind Flask and Streamlit REST interfaces.

  5. 2025

    Open-source contributor

    GirlScript Summer of Code '25 · Open Source Connect India

    Python contributions across community projects, before moving on to upstream fixes in Pydantic, NVIDIA NeMo, ModelScope, IBM and AWS Labs.

05 / Recognition

Stages won. Rooms entered.

Viksit Bharat Young Leaders Dialogue2026

State Rank 1, Rajasthan. National representative.

Top 0.0005% of 50.49 lakh aspirants. Represented Rajasthan at Bharat Mandapam, New Delhi, and presented Krishyak to the Union Agriculture Minister and the Director General of ICAR.

#1of Rajasthan
Chief Minister's Residence, Rajasthan2026

Invited to speak with the Chief Minister

One of a small group of youth invited to an interaction with Shri Bhajan Lal Sharma. My idea on taking rural products to national and global markets was cited by name.

Smart India Hackathon2026

Internal hackathon winners

Won BIT Jaipur's round with Krishyak and Team Stack Underflow, nominated for the national stage, while heading the committee that organised the 46-team event. SIH 2024 finalist.

IETE Engineers' Day2026

1st place, Student Project Competition

Krishyak took the top spot in Jaipur.

1st
NASA Space Apps Challenge2025

Global Finalist Nominee

For Intuix, an exoplanet detection system built on machine learning.

IETE Software Design Competition2025

1st prize

Top honours in the software design competition.

1st
Product HuntJul 2026

#11 Product of the Day

LLMSlim's first launch, where one builder reported 50% fewer tokens on their RAG pipeline with no quality drop.

#11
HackerRankPython

5-star Python

Plus 37 certifications in LLMOps, RLHF, post-training of LLMs, code agents and Google Cloud.

5★

06 / About

Building between classes.

I'm a BCA student at BIT Mesra's Jaipur campus with a minor in Artificial Intelligence. Most of my learning happens between coursework and whatever I'm shipping that week.

LLMSlim started with a question I kept running into: why are we sending so much context to models? Krishyak took the same engineering into fields and mandis. I like the whole path from research prototype to product: framing the problem, testing the model, building the interface and putting it in front of real people.

Today I split my time between LLMSlim, client work through AMEYOR, and fixing things upstream in the open-source tools I depend on. If you need a Python backend, a retrieval or agent system, an AI feature that actually works, or a fast, polished website, let's talk.

Education Birla Institute of Technology, Mesra · Jaipur Campus Bachelor of Computer Applications, minor in Artificial Intelligence · 2024 to 2027
9.35 / 10 CGPA

Languages

  • Python
  • TypeScript
  • JavaScript
  • C / C++
  • Java
  • SQL
  • Solidity

AI and ML

  • PyTorch
  • TensorFlow / Keras
  • scikit-learn
  • Hugging Face
  • Sentence Transformers
  • LightGBM
  • RAG
  • NLP
  • Computer vision

Agents and LLMs

  • MCP
  • OpenAI Agents SDK
  • LangChain
  • LlamaIndex
  • Haystack
  • Ollama
  • vLLM
  • Nemotron

Product and infra

  • FastAPI
  • Flask
  • Next.js
  • React
  • Postgres / PostGIS
  • Docker
  • Vercel
  • AWS
  • Linux
  • GNU Radio
Certifications and learning · 37
  • VBYLD 2026: Rajasthan State Rank 1 · Rajasthan Youth Board, Government of Rajasthan
  • VBYLD 2026: National Participation · Ministry of Youth Affairs & Sports, Government of India
  • Certificate of Completion | Breast Cancer Detection · Unstop
  • Python Programming Fundamentals · Microsoft
  • AI Essentials Google · Google
  • Building SnapAR Lenses with Lens Studio and BharatXR · Indian Institute of Technology, Delhi
  • Orchestrating Workflows for GenAI Applications · Astronomer
  • Postman API Student Expert · Postman
  • LLMOps · Google
  • Reinforcement Learning From Human Feedback · DeepLearning.AI
  • Post-training of LLMs · University of Washington
  • Building Code Agents with Hugging Face smolagents · DeepLearning.AI
  • Pydantic for LLM Workflows · DeepLearning.AI
  • AI Python · DeepLearning.AI
  • Problem Solving (DSA) · HackerRank
  • Advanced CSS · HackerRank
  • AWS APAC - Solutions Architecture Job Simulation · Forage
  • Mastercard - Cybersecurity Job Simulation · Forage
  • Electronic Arts - Software Engineering Job Simulation · Forage
  • JavaScript Portfolio Development using GitHub Codespaces (MLSA) · Microsoft Student Ambassadors - BIT Jaipur Chapter
  • Python Basics · United Latino Students Association
  • Basics of Figma · Simplilearn
  • Prompt Design in Vertex AI Skill Badge · Google
  • Analyze Images with the Cloud Vision API Skill Badge · Google
  • Networking Fundamentals on Google Cloud Skill Badge · Google
  • Monitoring in Google Cloud Skill Badge · Google
  • Get Started with Pub/Sub Skill Badge · Google
  • Cloud Speech API: 3 Ways Skill Badge · Google
  • App Engine: 3 Ways Skill Badge · Google
  • Cloud Functions: 3 Ways Skill Badge · Google
  • Get Started with Google Workspace Tools Skill Badge · Google
  • Get Started with Dataplex Skill Badge · Google
  • Get Started with Looker Skill Badge · Google
  • Get Started with API Gateway Skill Badge · Google
  • Get Started with Cloud Storage Skill Badge · Google
  • Develop GenAI Apps with Gemini and Streamlit Skill Badge · Google
  • The Basics of Google Cloud Compute Skill Badge · Google

07 / Contact

Have something difficult to build?

I take on freelance and client work through AMEYOR: AI features, RAG and agent systems, Python backends, and fast, polished websites. Tell me what you're trying to do and you'll get a considered reply, usually within two working days.

Freelance and client projects
hello@ameyor.in
Founder, direct
founder@llmslim.app
LLMSlim support
support@llmslim.app
Book a free 30-minute callTalk through scope, timeline and fit, no strings attached.
Pick a time
Find me elsewhereBuild logs, launches and demos.

Demo Open in a new tab  
Esc
↑ ↓ to moveEnter to select