This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle
What I Built
CircuitSathi is a local-first, privacy-preserving AI engineering companion designed for electronics students, robotics enthusiasts, makers, and small hardware prototyping teams.
The Problem & The "Side Hustle" Opportunity
When building embedded systems or diagnosing PCB designs, engineers consult hundreds of pages of component datasheets (pinouts, absolute maximum ratings, thermal dissipation limits, and timing diagrams). Today, using cloud AI for this has two major dealbreakers:
Intellectual Property & Privacy Risks: Hardware startups and consultants cannot upload proprietary circuit schematics, unreleased PCB designs, or client project documents to third-party cloud AI APIs.
Hallucination in Critical Systems: Generic chatbots frequently invent pin numbers or electrical specs and perform broken arithmetic on Bill of Materials (BOM) costs.
Who Would Pay For It?
Small Hardware Prototyping Firms & Repair Businesses: A one-time license or managed local installation for an air-gapped engineering copilot that keeps confidential client IP strictly on-premises.
University Robotics Clubs & Labs: Packaged offline toolkits for lab workstations where students can safely analyze motor drivers and power rails without metered cloud subscription bills.
Freelance Hardware Consultants: Automated, deterministic Bill of Materials (BOM) generation and styled Excel reports exported in seconds.
Demo
Demo
https://circuitsathi.streamlit.app/
Code
https://github.com/AadiptoGhosh/CircuitSathi
How I Built It
CircuitSathi is built around an entirely open, local-first stack with zero mandatory cloud dependencies:
Local Open-Weight Inference (Gemma 3 via Ollama):
Uses Gemma 3 (gemma3:4b as default, gemma3:1b for lower-VRAM laptops) running via a local Ollama server.
If Ollama is offline or the model is missing, CircuitSathi displays exact terminal commands (ollama pull gemma3:4b) instead of silently routing to a cloud vendor.
Page-Preserving Document Ingestion:
Built with PyMuPDF (pymupdf) to preserve 1-indexed page provenance for every text chunk.
Includes Tesseract OCR fallback for scanned technical diagrams and legacy datasheets.
Local Vector Search (ChromaDB + Sentence Transformers):
Embeds text locally using sentence-transformers/all-MiniLM-L6-v2 with unit-normalized embeddings.
Stored in a persistent local ChromaDB vector database with cosine distance matching.
Provides verified citations formatted as [Doc: , Page: ] and expandable evidence passage drawers.
Deterministic Engineering & Safety Engine:
Rather than letting a probabilistic LLM do math, Python handles all calculations:
Standard E12 resistor series selection & wattage derating for LEDs.
Unloaded voltage divider equations & power dissipation.
Deterministic Safety Checks: Automatically warns on SELV shock hazards, missing inductive flyback diodes, logic-level mismatches (5V TTL into 3.3V ESP32 GPIOs), and lithium battery BMS requirements.
Validated BOM Accounting & Export:
Pydantic v2 validation for line items and quantities.
Python-calculated financial subtotals with unpriced item flags.
One-click export to professional, styled Excel workbooks (.xlsx) via OpenPyXL and CSVs via Pandas.
Why Does Open Innovation Matter?
Hardware engineering demands trust, safety, and confidentiality:
True Intellectual Property Protection: Hardware designs and pre-release schematics are core business assets. Open innovation and open-weight models like Gemma 3 make it possible to run state-of-the-art AI locally on a developer’s laptop without telemetry or data harvesting.
No Recurring Per-Token Tax: Hobbyists and student teams prototyping robotics shouldn't have to budget per-token API calls or risk getting locked out during a competition because their API credit ran out.
Reproducibility & Auditability: Because the weights, embedding models, and vector stores are open, every retrieved passage and citation can be audited against the manufacturer's original PDF.
My Agent Session
This project was architected, implemented, and benchmarked (passing 21 automated unit tests and a 20-question datasheet benchmark with a 90% retrieval hit rate and 100% citation accuracy) using DevRelay.
My First User
I shared an early build of CircuitSathi with a student member of our university robotics club who was debugging a dual-motor differential drive robot that was failing to turn:
"Normally I have to dig through a 25-page Toshiba PDF to figure out why my TB6612FNG driver wasn't responding. CircuitSathi immediately pointed out that the STBY (Standby) pin was measuring 0V and had to be pulled HIGH to VCC, citing Page 2 of the datasheet. The deterministic safety warning also reminded us that we hadn't wired common ground between the 7.4V battery pack and the microcontroller. That 5-minute diagnosis saved us from burning another motor driver."
They asked if they could install it on the club's offline benchtop lab laptops for upcoming competition prep.
Top comments (0)