Applied AI Systems Engineer · Founder, QuantaSharp · Pune, India
I build AI systems that check their own work.
I'm Aaryan Patwardhan. The AI does the reading, fixed rules check the result, and anything uncertain goes to a person — tested, measured and shipped to real users.
- AI model
- Deterministic rule
- A person reviews
- Not working yet
- inputDocuments and data
- AI modelAI reads and proposes
- ruleRules verify the result
- personDoubts go to a person
- outputSaved, sent or acted on
Selected work
Each project with its real pipeline, its honest stage, and only numbers I've measured.
- Pilot-ready · v3.4Automation · Backend · Vision AI
InvoiceAI
Reads Indian GST invoices — digital PDFs, scans and phone photos — checks its own work, and marks anything doubtful for a person to review.
- 83% field accuracy, 0 invented values (11 sample docs)
- 136 automated tests
- Accuracy gate in CI
- Commercially licensed models only
- inputPDF, scan or photo
- AI modelLocal LLM reads it
- ruleRules check GSTIN, totals, tax
- personDoubts go to a person
- storageSaved · Excel export
Case study: InvoiceAIPrivate code — demo on request - LiveFull-Stack · Backend
CramIQ
Exam preparation for SPPU engineering students: a study planner, a study vault, progress streaks and a daily briefing email.
- Live for students
- Sandboxed service (systemd exposure 1.3)
- Deploys roll back on failure
- Nightly briefing emails
- inputNext.js app
- serviceFastAPI, sandboxed VM
- storagePostgreSQL + Redis
- AI modelGemini study material
- outputDaily briefing email
- Hackathon · team leadAutomation · Backend · Systems
Hyron AI
Finds the risky packages in a project's npm dependencies, explains each score, and plans safe upgrades. FUSION 2026 hackathon — team lead.
- 24-hour hackathon build
- OSV + EPSS + CISA KEV
- Explainable 0–100 score
- Team of four
- inputnpm lockfiles
- serviceDependency graph
- ruleOSV + EPSS + CISA KEV
- ruleExplainable risk score
- personUpgrade plan: auto / review / block
- PrototypeSystems · Backend · Full-Stack
Aura — attendance from campus Wi-Fi
Turns Wi-Fi accounting events into attendance records — no student app. A working prototype driven by a simulator.
- Prototype
- Redis Streams
- PostgreSQL
- Docker Compose
- inputWi-Fi events (simulator)
- serviceFastAPI ingestion
- storageRedis Streams
- AI modelAttendance rule + Isolation Forest
- outputPostgreSQL + dashboard
- Hackathon · AI core leadVision AI · Automation · Systems
SentinelMesh
AI-assisted drone response: YOLOv8 detection with two local LLM agents arguing before a dispatch. INSPIRON 5.0 (COEP) — AI core lead.
- Hackathon build
- YOLOv8
- Two-agent debate
- Simulated fleet
- inputRecorded video
- AI modelYOLOv8 detection
- ruleConfidence gate
- AI modelTwo LLM agents debate
- outputDispatch on a simulated fleet
Case study: SentinelMeshPrivate code — demo on request - Prototype · in progressSystems · Automation
Ghost-Admin (prototype)
A Linux daemon designed to notice runaway processes and step through safe, logged fixes. Work in progress — the AI step is not working yet.
- Prototype
- 7-stage design
- JSONL audit log
- Local LLM via Ollama
- inputRAM trend detection
- ruleWhitelist + baseline
- AI model · not working yetLocal LLM classifies
- outputGraduated remediation
- storageJSONL audit log
How I build
Four rules I hold every system to. You can check each one in the projects above.
Measure before claiming
Accuracy comes from a labelled test set and a script anyone can re-run — never from a guess.
Let rules check the AI
The model proposes; deterministic checks verify. GSTIN check digits, totals, confidence gates.
Show every doubt
Anything uncertain goes to a person, in plain words. Nothing doubtful is marked complete.
Ship it like production
Tests, CI, Docker, sandboxed services and rollbacks — on a cloud VM or a client's own machine.
Experience
QuantaSharp, Pune
Apr 2026 – Present
Founder & Applied AI Systems Engineer
- Built InvoiceAI, which reads Indian GST invoices (digital PDFs, scans and phone photos) with local LLMs via Ollama behind a FastAPI service with SQLite storage.
- Designed a deterministic verification layer: GSTIN check-digit validation with OCR repair that is accepted only when the checksum passes, totals and CGST/SGST cross-checks, and seller/buyer correction from PDF text positions. Every doubtful field is flagged for review.
- Wrote 136 automated tests and a labelled golden-set accuracy gate in GitHub Actions, so a model or prompt change that lowers accuracy cannot be merged; 83% field accuracy with 0 invented values on 11 sample documents.
- Shipped it as an offline Docker package with Windows and Linux installers, and designed the dashboard for non-technical and older users.
Code N Try, Pune
Jan 2026 – Feb 2026
Python AI/ML Development Intern
- Built a house-price prediction model: explored the data with pandas and Seaborn, trained a scikit-learn regression on a train/test split, and saved it with joblib for a prediction script.
- Built a Flask and SQLAlchemy web store with a product catalogue, session cart, checkout and an admin dashboard for adding and editing products.
- Built a command-line chat assistant on the Google Gemini API, with the API key kept out of the code.
Education
B.E. Information Technology
SKN Sinhgad Institute of Technology & Science, Lonavala (Savitribai Phule Pune University)
Expected Jun 2028
Hackathons
- FUSION 2026 — team lead, Hyron AI (participant)
- INSPIRON 5.0, COEP — AI core lead, SentinelMesh (participant)
- Smart India Hackathon 2026 (participant)
- Decode SIH 2026 (participant)
Certifications
- Front-End Development Internship — Cognifyz Technologies (2026)
- 21 Projects, 21 Days: ML, Deep Learning & GenAI — GeeksforGeeks (2026)
- Python Skill Up — GeeksforGeeks (2026)
Stack
What I've used in the projects on this page — and what I'm learning next.
AI and LLMs
- Local LLMs via Ollama
- Vision-language models for documents
- Gemini API
- LLM evaluation with golden sets
- YOLOv8
- scikit-learn
Backend and data
- Python
- FastAPI
- PostgreSQL
- Redis / Redis Streams
- SQLite
- SQLAlchemy
Infrastructure and delivery
- Docker and Compose
- GitHub Actions CI
- Linux servers, systemd sandboxing
- nginx
- Vercel
Frontend
- TypeScript
- React
- Next.js
- Tailwind CSS
Learning now
From my six-month plan — being learned, not claimed.
- Agentic AI: tool calling, MCP, LangGraph
- AWS (Solutions Architect Associate)
- Terraform
- Kubernetes, GitOps, observability
- Go
- MLOps: MLflow, model serving
Let's build something that works.
Open to applied AI and backend internships, and to QuantaSharp projects. I reply within 24 hours.

