Project 001 · Pharmacy retail · Document AI
Distributor bills into checked stock, from a photo.
99.5%of held-out bill line values read right, at $0.03 a bill.
React · Express · MongoDB · FastAPI · OpenAI vision
Sarthak Aggarwal · AI engineer · India
AI systems that hold up in production.
RAG, multi-agent systems, document intelligence, voice AI, MCP and LLM security, built end to end: interface, backend, models, database and deployment. Every number I show is measured on data the system never saw.
Shamli, India · IST
- AIR 1992GATE 2024, Data Science & AI
- Top 5Smart India Hackathon 2024
- ExpertKaggle Notebooks
- Top 3%Amazon ML Summer School
- 2 patentsPublished
What I build
Six areas,one engineer end to end.
01
RAG & hybrid searchKeyword + vector search, reranking and cited answers that say so when the sources don't cover a question.02
Multi-agent systemsLangGraph supervisors, tool-using agents and a verification agent before anything reaches a person.03
Document intelligenceOCR and vision models on scans, photos and PDFs; every extracted number is checked in code.04
Voice AISpeech-to-text, text-to-speech and calling agents, chosen on latency, quality and cost.05
MCP & tool useModels wired to real tools and data through MCP servers, each with a narrow, explicit scope.06
LLM securityPrompt-injection defences, read-only SQL gates, output checks and least-privilege access.
Approach · 2026
Most AI demos impress. Few survive production.
01
Real data
Messy documents, real regulations, live calls, not clean sample text.
02
Measured
Held-out evals, labelled with version and sample size.
03
Full stack
React, Node, Python, queues, Postgres, Docker. One person, no hand-offs.
04
Checked
The model reads. Python does the maths. A person signs off.
I don't ship prototypes
systems.
- Evals before claims
- Every answer traced to its source
- Runs on one small server
Selected work.Each one measured.
3 projects · each with a held-out eval
Project 002 · Lending · Document AI
Loan income verification that shows its working.
11 / 11planted fraud cases caught, with no honest applicant flagged.
React · Node · RabbitMQ · Postgres · Mistral OCR · OpenAI
Project 003 · Lending compliance · RAG
RBI rules answered, with the paragraph to prove it.
94%of held-out questions had the right paragraph in the top five.
Hybrid search · Rerank · Text-to-SQL · FastAPI · Postgres
How I build
Four rules,in every project.
01LLM + Python
The model reads. Python checks.
LLMs copy values and write text; sums, rules and verdicts are plain code you can test.
02Held-out evals
Measured on unseen data.
Every number is from a held-out set written after the last change, labelled with version and size.
03React · Node · Python
Full stack, one owner.
Interface, API, queue, model service, database and Docker, built and run by one person.
04Citations
Every claim traceable.
Each answer, flag or amount points back to the page, paragraph or row it came from.
Experience
Data Scientist
Tuning Research Private Limited (WorkableAI)
Feb 2026 – now
Own the AI side of client products end to end: multi-agent LLM pipelines with hybrid RAG for export compliance, a web-extraction pipeline into a 40-table database, and OCR → extraction → LLM-judge workers for question banks.
Business Analyst Intern
Tuning Research Private Limited (WorkableAI)
Jul 2025 – Feb 2026
Started as BA, took over product, AI architecture and development, then converted to full time.
SDE Intern
Cruv
Jul – Sep 2025
WhatsApp and AI voice-calling agents for insurance renewals; benchmarked TTS models on quality, latency and cost.
Achievements
01 · 2024
GATE DA, AIR 1992Data Science & Artificial Intelligence paper.02 · 2024
SIH Top 5 finalistSignify: Indian Sign Language to text and speech.03 · 2024
Amazon ML Summer SchoolSelected, top 3% of applicants.04
Kaggle Notebooks ExpertMultiple high-ranking ML notebooks.05
2 patents publishedEarly forest-fire prediction; Signify.06 · 2025
99.45% on MalimgImage-based malware classification research.
Education
B.Tech, Computer Science
KIET Group of Institutions · CGPA 8.3 · Minor in ML
2022 – 2026
Core Lead, School of AI, ML & DS: ran 10+ events and hackathons with a 15-person team.
Class XII · Class X
CBSE · 95.6% · 96.2%
2021 · 2019
Writing.Notes on AI systems.
All posts on Medium ↗Contact
Say hello.I read everything.
Questions about a project, an idea to talk through, or just AI shop talk.