Gautam Singh
Software engineer. Backend, data and ML systems.
Computer Science honours grad, currently a machine learning intern on an
automated vehicle damage inspection system. I build data pipelines, ML
systems, and the full-stack products on top of them. Three of my apps are
live in production, one on its own domain with a published
MCP server on npm.
Resume (PDF)
gautamsingh2189@gmail.com
GitHub
LinkedIn
Open to work. Guelph, Ontario, and open to Kitchener-Waterloo, Toronto, or remote. I answer every message.
- Live in production
- 3
- Projects shipped
- 10
- Co-op and internship terms
- 3
- BCS Honours, Conestoga
- 2025
Selected work
Pick an upload limit and get a file that fits under it. PDFs and images, iPhone HEIC photos included, on its own domain with its own npm package.
- Shipped seven file formats, 37 limit-specific landing pages and 12 form presets for visa and exam uploads
- Each preset sets the size cap, the minimum size and the exact pixel dimensions the form demands
- Cut a 5 MB photo to 200 KB from 9.2 s to 4.2 s in production by guessing the next setting from sizes already measured
- Halved renders per run from four to two, with every output byte-identical
- Cut the wait after Compress from 3.9 s to 0.3 s by starting the run while the visitor reads the size picker
- Resampled large photos in the browser, so a 10.7 MB upload goes up as 1.5 MB, 8.9 s instead of 57.7 s
- Verified browser resampling reaches the same answer in 23 of 24 photo-and-limit pairs
- Published an MCP server to npm, listed in the official Model Context Protocol registry
- Runs the API and two workers in Docker Compose on a VPS behind a Cloudflare Tunnel, React frontend on Vercel
- Added a watchdog that restarts any container the health endpoint reports stuck
- Autodeploy ships main within two minutes of a merge
PythonFastAPIReactTypeScriptRedisRQDockerGhostscriptPillowCloudflareVercelMCP
Live app
Code
npm package
Self-hosted web analytics. Rust on Lambda, no cookies, no IP addresses stored, and it is counting this page view.
- Ingest is one DynamoDB write per event, with bot filtering, user-agent classification and URL normalization as pure CPU
- Measured 184 ms cold start, 5.1 ms warm and 30 MB peak on arm64 at 128 MB
- Hashed visitor IDs with keyed BLAKE3 under a daily salt, so IDs cannot be correlated across days
- The persisted type has no IP field, and query strings are stripped before write
- Tracker is 647 bytes gzipped, wraps pushState for SPA routes and writes nothing to the browser
- Recomputes whole days from raw rows rather than incrementing counters, so re-runs are idempotent
- Deployed the stack as 41 Terraform resources, $0/month on perpetual free tiers
RustAWS LambdaDynamoDBCloudFrontAPI GatewayS3Terraform
Live dashboard
Code
Invoices and receipts into structured data, with per-field confidence and a human review queue. Runs locally on one consumer GPU.
- Raised field accuracy from 0.804 F1 to 96.7% by auto-accepting only trusted fields and routing the rest to a person
- Scored confidence as self-consistency agreement across repeated runs, capped when arithmetic or format validation fails
- Trusted fields run 94.4% accurate against 54.1% for flagged ones
- A subtotal equal to the total with no tax printed goes to a human instead of into the books
- Wrote 175 tests, most from real failures found in error analysis
- CI re-scores frozen predictions on every push, so a silently moved number fails the build
- Measured calibration at ECE 0.115 with a reliability table, rather than assuming it
- Runs locally at $0.00 per document against roughly $0.004 on a hosted vision API
PythonPyTorchQwen2.5-VLFastAPIPydanticGitHub ActionsDocker
Cloud data pipeline and forecasting platform, built with my capstone team at Conestoga.
- Sped up data loads ~35% by redesigning the storage layer of the Databricks and Azure Data Factory pipeline
- Trained a 2-layer LSTM on 11,213 sliding windows to predict next-week returns, tested with walk-forward validation
- Used a Django REST API to send forecasts to a web client, a Flutter app and Power BI
PyTorchDatabricksAzure Data FactoryPySparkAzure SQLDjango RESTFlutterGrafana
Team repo
Upload a CSV and get back type inference, quality flags, generated insights, and a chart for every column.
- Cut the chart bundle from 5.1 MB to 1.8 MB by trimming the Plotly build
- Validated uploads to CSVs under 20 MB with structured error responses
- Reports expire after five days through a background cleanup job
FastAPIReactSupabasePlotlyVercelRender
Live app
Also shipped
Fixed a silent failure in a local-first finance app with 29k GitHub stars, where changing a password on an account with no password method set returned nothing. Reproduced the open issue, wrote the fix and took it through review.
TypeScriptOpen sourceCode review
Answers plain-English questions over real SEC 10-K filings. BM25 and FAISS run together with reciprocal rank fusion, which beat either retriever alone, and a local Gemma model streams the answer. Retrieval scored with RAGAS.
LangChainFAISSBM25RAGASDocker
Daily pipeline pulling OHLCV data for S&P 500 stocks and sector ETFs into PostgreSQL. dbt builds staging tables, technical indicators, and rolling performance metrics. Prefect runs it at market close and feeds a Streamlit dashboard.
PostgreSQLdbtPrefectStreamlit
Fixed-income interest-rate risk framework. Computes DV01 across the curve from 3M to 30Y, runs stress scenarios including parallel shifts, bear steepeners, and bull flatteners, and attributes PnL by maturity bucket.
PythonPandasNumPyFixed income
Classifier predicting next-day price direction for AMD. Pulls history from Yahoo Finance, builds moving averages, RSI, and MACD with the ta library, and trains a HistGradientBoostingClassifier evaluated on a held-out test set.
scikit-learnPandasta
Interactive lessons on the data structures and algorithms that come up in coding interviews, with visual examples that run in the browser. Built as a study tool.
JavaScriptAlgorithms
Experience
Machine Learning Intern
Sphere Global · Remote
Aug 2026 to present
- Built a generator that turns clean vehicle photos into YOLO-labeled damage with pixel-exact masks, ~1 s per image on CPU
- Produced around 12,000 labeled images on one CPU-only laptop, keeping customer photos on-site
- Benchmarked SDXL and FLUX against a procedural optical model, which won on accuracy, speed and cost per image
- Root-caused a labeling bug that mislabeled 39% of training images; added a YOLO vehicle check before SAM 2.1 segmentation
- Handed the labeling team a single 84 MB executable, so teammates without Python could run it
- Labeling tabs write the team’s existing 79-class Label Studio format
ServiceNow Developer & QA Specialist
PAMT Consulting, now Ateko (a Bell Canada company) · Remote
Sep 2024 to Dec 2024Sep 2025 to Dec 2025
- Cut new-hire onboarding handling time ~30% by replacing a manual process with a Flow Designer workflow
- Wrote and ran test cases for workflows, integrations and APIs across dev, test and production
- Tracked down bugs in client builds and checked each fix in every environment before release
- Kept update sets organized and wrote up test results for the dev team
Lot Associate
The Home Depot · Guelph, Ontario
Jan 2022 to May 2022
- Kept the lot stocked and organized and helped customers in a high-traffic store
Tools
- Languages
- PythonSQLJavaJavaScriptTypeScriptC++CRust
- Data engineering
- Azure Data FactoryDatabricksPySparkSparkHadoopdbtPrefectETL/ELT
- Machine learning
- PyTorchTensorFlowXGBoostLightGBMscikit-learnOpenCVYOLOLangChainFAISSLSTM
- Databases
- PostgreSQLMySQLMongoDBAzure SQLDynamoDBRedisSupabase
- Web and tools
- DjangoFastAPIReactFlutterStreamlitMCPServiceNowJMeter
- Cloud and DevOps
- AzureAWSAWS LambdaTerraformDockerCI/CDVercelCloudflareGitGrafanaPower BITableau
Education
Bachelor of Computer Science (Honours)
Conestoga College
Sep 2020 to Dec 2025
- Machine learning, data structures and algorithms, database design, cloud computing, and software engineering
- Capstone: The Almanac, a cloud pipeline and forecasting platform
Google Advanced Data Analytics Professional Certificate
Google · Coursera
Jun 2026
- Seven-course program covering statistics, Python, regression analysis, machine learning, and visualization