LLM systems, end to end
From a fuzzy idea to a deployed product: model selection, orchestration across GPT / Claude / Gemini / LLaMA, prompt architecture, evals, and the boring reliability work that decides whether it ships.
Open to work: freelance & full-time
Hey, I'm Jay Thakkar.
I build LLM systems that survive production: MCP infrastructure, voice AI agents, and RAG pipelines that stay fast, cheap and honest under real traffic. 3+ years shipping for enterprise contact centres, fintech and SaaS teams. Google Developer Expert for AI & Cloud.
JT
🎙️ Voice AI
🔌 MCP
📚 RAG
- 0%Query latency cut
4s → 300ms via MCP - 0%LLM compute spend
reduced - 0%Hallucinations cut
with prompt guardrails - 0Developers trained
APAC · MENAT · SSA
Freelance & consulting
What I can build for you
Scoped engagements delivered end to end, covering architecture, implementation, evals and handover. Five enterprise clients shipped, 95% satisfaction, integration timelines cut by 40%.
Voice AI agents
Contact-centre grade voice agents: intent design, prompt tuning, barge-in and latency budgets, system-level guardrails. Improved response reliability by 15% and cut hallucinations 87.5% on live deployments.
MCP integrations
Model Context Protocol servers that connect your models to real systems: 10+ data sources, VM-based auth handling 500+ secure API transactions a day, zero downtime. Query time down 92.5%.
RAG pipelines
Ingestion, chunking, hybrid retrieval and re-ranking that actually improves answers. I've taken retrieval accuracy from 30% to 90% and lifted generation quality 40% on client corpora.
AI cost optimisation
Your bill is usually a design problem. Caching, model tiering, context discipline and retrieval hygiene, for a 60% inference cost reduction without giving up quality.
Advisory, workshops & talks
Google Developer Expert for AI & Cloud. Team enablement, architecture reviews, and hands-on workshops, using the same material that's reached 30,000+ developers across three regions.
How engagements work
Discovery sprint: one to two weeks, ending with an architecture and a costed plan. Build: fixed-scope delivery with weekly demos. Retainer: ongoing tuning, evals and on-call for the system once it's live.
Experience
Where I've built
Impulsive Web → Technology 9 Labs
Jun 2025 - PresentFull-time · Hybrid · Impulsive Web acquired by Technology 9 Labs, 2025
Resident Entrepreneur, Technology / AI EngineerDec 2025 - Present
- Promoted from AI Engineer to Resident Entrepreneur on acquisition, and now own AI infrastructure strategy and enterprise client delivery.
- Architected MCP infrastructure across 10+ data sources, cutting query response time 92.5% (4s → 300ms); VM-based auth processing 500+ secure API transactions daily with zero downtime.
- Tuned voice AI prompts with Chain-of-Thought and few-shot techniques, giving 15% better LLM response reliability across enterprise contact-centre deployments for 3CLogic.
- Shipped system-level prompt guardrails cutting hallucinations 87.5%, and a CLI tool bridging backend APIs with Claude and Codex that took solution cycles from 15 weeks to 2.
- Full-cycle AI integration for 5+ enterprise clients via custom MCP pipelines: 40% faster data onboarding, 95% client satisfaction.
AI Engineer, Impulsive WebJun 2025 - Nov 2025
- Scoped and architected end-to-end MCP-based AI pipelines for enterprises across fintech, logistics and SaaS, cutting integration timelines by 40%.
- Led cross-functional sprints shipping production-ready AI modules on schedule, NPS above 95% across engagements.
Forsyt.ai: Revenue Operating System
- Built the AI Revenue Intelligence engine (Python, React, PostgreSQL, Gemini) ingesting CRM, email and calendar data to surface risk and auto-generate action bundles, lifting deal adherence 15%.
- Action Flow Engine for one-click AI-drafted email and CRM updates, saving 10+ hours per rep per week.
- Confidence Scoring model on sentiment and engagement velocity, improving prediction of deal outcomes by 25%.
Hoodo: product startup
- Launched an intelligent calendar platform automating 100+ meetings per participant.
- AI meeting-notes system with full audio capture and structured note generation, zero content loss.
SCAAI (Symbiosis Centre for Applied AI)
Jul 2023 - PresentPart-time · 2 yrs 7 mos
R&D Mentor, Gen AIJun 2024 - Present · Pune, Maharashtra · Hybrid
Research InternJul 2023 - Jul 2024 · Remote
Researcher
Jan 2025 - May 2025Truxt.ai · Full-time · Remote
Binoloop
Jan 2024 - Jan 2025Canada · Remote · 1 yr 1 mo
Machine Learning AssociateOct 2024 - Jan 2025 · Full-time
AI/ML Research AnalystJan 2024 - Oct 2024 · Internship
- Promoted from research analyst to associate within the same tenure for RAG pipeline and multi-LLM orchestration work.
- Built and deployed RAG pipelines with advanced text extraction, giving 40% better generation quality at 60% lower inference cost.
- Orchestrated solutions across Gemini, GPT, Claude and LLaMA, a 30% uplift in accuracy and efficiency across 8+ client engagements.
AI Intern
Nov 2023 - Feb 2024Otobit Private Limited · Full-time · Surat, Gujarat · On-site
Earlier roles, 2020 to 2023
-
Student CoordinatorPPSU Data Science Club · Full-time · Surat, GujaratAug 2022 - Sep 2023
-
Project ManagerTriclone Technologies Private Limited · Full-time · Surat, Gujarat · On-siteDec 2022 - May 2023
-
Teaching InternP P Savani University · Part-time · Surat, Gujarat · On-siteOct 2022 - Apr 2023
-
Business Intelligence SpecialistPGP Glass Pvt. Ltd. · Internship · Vadodara, GujaratJun 2022 - Jul 2022
-
Data Scientist → Data Analyst → Software Developer TOPS Technologies Pvt. Ltd · Internship · Surat, Gujarat · Data Scientist (Oct 2021 - Aug 2022), Data Analyst (Oct 2021 - Apr 2022), Software Developer (Jun 2021 - Oct 2021)Jun 2021 - Aug 2022
-
Software DeveloperTOPS Technologies Pvt. Ltd · Internship · Surat, Gujarat · Java developer who built a Java-based websiteNov 2020 - Mar 2021
Toolkit
Languages
ML & LLM
AI platforms
Data
Cloud & DevOps
TPU Sprint Q1 2026
vForge
Benchmarks LLM fine-tuning on TPU against GPU end to end: a chat-driven dataset builder, LoRA fine-tunes in JAX and Keras 3, then throughput and cost compared with vLLM.
View repo
★ 12
4 forks
24K views
AI Hiring Agent
Five specialist agents under one conversational orchestrator, scoring resumes and validating GitHub profiles live before returning a hiring verdict.
View repo
★ 22
Most starred
50K views
RAG with Gemini
Retrieval-augmented PDF analysis on Gemini, worth a 5% lift in question-answering accuracy over the baseline. The most-read thing I have published.
View repo
Codelab
Multi-Agent Coding Assistant
Planner, coder and explainer agents collaborating in real time, each streaming its own live UI panel into a Next.js app over the 2026 agent protocol stack.
View repo
Fully offline
Local RAG with Gemma
Retrieval-augmented answers where no document leaves the machine: Gemma 3 through Ollama, a FastAPI backend and a numpy vector store with no external database.
View repo
★ 5
3 forks
Task Prioritisation Agent
Ranks a task list by urgency, importance, effort and company OKRs with Vertex AI Gemini, served as both a REST API and a small containerised web app.
View repo
Community
Google Developer Expert for AI & Cloud
Since March 2024 I've been teaching AI/ML in the open through Google-backed programmes, community talks and hands-on workshops that have reached 30,000+ developers.
- Drove AI literacy across APAC, MENAT and SSA through training programmes and technical workshops.
- Grew reach from APAC-only in 2024 to three regions inside 12 months, tripling geographic coverage and speaking engagements year over year.
Education
M.Tech in Artificial Intelligence & Machine Learning
Symbiosis Institute of Technology
Aug 2024 - May 2026
B.Tech in Computer Science Engineering (AI & ML)
P. P. Savani University
Oct 2020 - Jun 2024
Based in Surat, Gujarat, India
Working with teams across IST, CET, EST and PST
Research
Peer-reviewed publications
-
Evaluating Large Language Models for Knowledge-aware Question and Answering
International Journal on Smart Sensing and Intelligent Systems
-
Optimizing Energy Management in Smart Grids: A Hybrid Approach to Load Forecasting
PACIS 2025
-
A Pragmatic Approach on Adoption of EDA to Make Intelligent Business Decisions
International Journal of Wireless Network Security
Open to work: freelance & full-time
Let's build something that ships.
Have an LLM idea stuck at the demo stage, a voice agent that hallucinates, or a bill that keeps climbing? Tell me what you're working on. I reply to everything.