Vatsa Joshi.
building agents that ship.
Full-stack engineer working at the seam between large language models, multi-agent systems, and the messy production infrastructure that has to carry them. Currently shipping legal AI at Promact.
The human
behind the agents.
I build end-to-endsoftware — from low-level model implementations to production full-stack systems — with a founder's bias for shipping.
Computer science undergraduate at Navrachana University, graduating September 2026. I've spent the last three years working across the AI/ML and SaaS industries as an intern, contractor, and now a full-time trainee — specializing in agentic systems, retrieval pipelines, and the boring-but-critical infrastructure that makes LLMs behave in production.
The stuff I actually care about: multi-agent architectures with real coordination logic, RAG systems that don't hallucinate, fine-tuning small models for specific domains, and writing transformers from scratch in Rust when nobody asked me to.
Navrachana University · 2022–2026
Between builds. Looking
for what's next.
next build.
Just wrapped my tenure at Promact contributing to BharatLaw.ai — designing AI systems that read, parse, and draft across millions of Indian legal documents. Now open to full-time AI engineering roles where I can keep shipping agentic systems and production-grade LLM infra.
- Built end-to-end AI-driven Document Management Systems — translation, parsing, indexing, retrieval across 22+ languages.
- Cut legal NER extraction cost 25× — from ₹1/doc to 6-7 paisa — across an 80-lakh Indian court judgment corpus. Replaced a 7-stage LLM pipeline with a 2-call Gemini-2.5-Flash-Lite architecture.
- Shipped a multi-agent Petition Writer drafting structured legal petitions from case inputs, with iterative human-in-the-loop clarification.
- Designed a legal intelligence platform — judge profiling, case-brief-driven litigation strategy, lawyer compatibility analysis — agents coordinating vector similarity, LLM reasoning, and live web search.
A short trail
of roles.
Current
6 months
3 months
3 months
3 months
3 months
Things I've built
on purpose.
Tools of the trade.
Languages
- Python
- TypeScript
- Rust
- Java
- C
AI / ML
- PyTorch
- TensorFlow
- HuggingFace
- Unsloth
- Scikit-learn
LLM & Agents
- LangChain
- LangGraph
- CrewAI
- RAG / CAG
- LoRA · QLoRA
- FAISS · Groq
Backend & Cloud
- FastAPI
- NestJS
- Node.js
- Docker
- AWS
- PostgreSQL · Mongo · ES
Frontend
- Next.js
- React
- React Native
- Tailwind
Vision
- OpenCV
- YOLO / SSD
- PaddleOCR
- Diffusers
Data & Infra
- Elasticsearch
- Qdrant
- Upstash Redis
- Inngest
Practices
- End-to-end ownership
- Human-in-the-loop
- Shipping > perfecting
Honours & paper trail.
Tinkerthon 3.0
Second place for a Deepfake Video Detection System. Fine-tuned DL pipeline selected among top hackathon submissions.
SSIP Gujarat Grant
One-year research grant from SSIP Gujarat to further develop the deepfake detection project into a production-ready tool.
Odoo × NMIT Hackathon
Top 5 finish out of 3,300 competing teams — rapid product design and implementation against a shared brief.
B.Tech · Computer Science
Navrachana University, Vadodara. Specialisation in Software Engineering & AI. Research on quantum-dot memory, EfficientNetB7 + BiLSTM for deepfake detection, and stable-diffusion VAEs.
Rabin-Karp Algorithm
Report & analysis of the Rabin-Karp string-matching algorithm. Published on ResearchGate.
DOI: 10.13140/RG.2.2.24592.01283
Blockchain Technology
Survey paper exploring blockchain architecture, cryptographic foundations, and cross-industry applications.
DOI: 10.13140/RG.2.2.32465.75365
EfficientNetB7 & Bi-LSTM for Deepfake Detection ↗
Hybrid spatial-temporal pipeline for deepfake video classification — EfficientNetB7 frame features fed into a Bi-LSTM for sequence-level reasoning. First author.
Self-Adaptive Context RAG
Invited talk on a RAG architecture that tunes its retrieval context on the fly — delivered to the Google Developers Group community.
Agentic AI
Gave a workshop on multi-agent systems, tool-use patterns, and shipping agentic AI into production. Navrachana University.