A gamified interview prep prototype.
Selected work
Case studies
Supplier risk from geopolitical news, scored and explained.
- Top TalentManhattan Associates, 2025
- Microsoft for StartupsRolePrep.ai
- Google Cloud ML EngineerCertification earned in 2023
Jun to Aug 2026
Forecast Explainability Agent
Graduate Research Intern · Manhattan Associates, Atlanta
A summer internship during my Master's Degree, after five years at Manhattan.
- Forecast Explainability Agent: an agent with LLM tool calling that explains forecast changes in plain language, validated against a curated evaluation set.
- Rich visual output: SHAP contributions and waterfall charts rendered alongside the agent's explanation.
- Customer segmentation: UMAP and SHAP analysis, plus weighted scoring that categorizes weather and promotion data automatically.
Explain a forecast
Choose a scenario and open a driver to see what changed.
A spring promotion meets a competitor's price cut.
Source check All four sources updated within 24 hours.
Baseline100Total
- Buy-one-get-one offer+11
- End-cap display+5
- Markdown depth+2
- Warm weekend+4
- Dry days+2
- Competitor price cut−7
- List price change−2
- Easter proximity+3
- Payday week+1
Forecast119Total
Axis starts at 80. Break marks show the omitted range.
The forecast is 119 units, 19 above baseline. Promotions add 18, led by the buy-one-get-one offer. Weather adds 6, while Easter and payday timing add 4. Pricing offsets those gains by 9, mainly from a competitor's price cut.
Interactive illustration with fictional data and tool calls.
2025 to Dec 2026
Building while learning
Master's Degree in Engineering Management · Northwestern University
McCormick School of Engineering and Kellogg School of Management.
- RolePrep.ai, from first commit to a working prototype: interview prep made fun rather than boring, with points, flash cards and follow-ups built on a resume gap analysis and mock interviews. Selected for Microsoft for Startups.Self-started
- RolePrep's analysis in about 31 seconds, down from 90 to 150: moved the inference off the request path and co-located the function with the database, then added provider failover and a remote MCP server so Claude, ChatGPT and Cursor can run the same analysis.
- GeoSupply Copilot, the capstone: a supplier risk dashboard and retrieval-based assistant that bring order data and geopolitical news together.
- Coursework: Analytics, Operations and Supply Chain Management at McCormick, with Product Management at Kellogg.
How I built RolePrep.ai
- Make the analysis dependable
- Timing logs traced the delay to deployment behavior and database distance. Post-response execution and a nearby database brought the wait down to about 31 seconds.
- Plan for provider failures
- After Gemini outages, I added two fallback providers and retries. I tested each path by forcing the primary provider to fail and comparing the reports.
- Turn findings into practice
- Drill cards and mock interviews use the gaps already found in the analysis. Candidates can keep preparing without entering their background again.
RolePrep.ai
Prototype · selected for Microsoft for StartupsGamified interview prep: points, flash cards and follow-ups on top of a resume vs job-description gap analysis, with an MCP server for Claude, Gemini, ChatGPT and Cursor.
GeoSupply Copilot
Northwestern capstone · Publication in progressBrings geopolitical news and order data together to assess supplier risk. A retrieval-based assistant helps explain which orders may be affected.
Jul 2024 to Aug 2025
Senior Software Engineer, ML
Manhattan Associates
- 5x faster critical ML computations: Modin on Ray integrated into the platform while keeping existing workflows compatible.
- Proof of concept to production: owned the rollout of the scheduled notebook platform that data scientists run their experiments on, aligning the Data Science and Cloud Platform teams on scope and delivery.
- Hundreds of millions of records: kept the forecasting workflows reliable at that scale by adding fault tolerance to distributed Modin on Ray processing, and triaged production issues to keep them running.
Feb 2021 to Jun 2024
Software Engineer, ML
Manhattan Associates
- Forecasting platform from inception to production: helped build Manhattan's ML platform, with XGBoost and LightGBM models tuned and validated on holdouts.
- 15x faster PySpark stages: tuned the compute-heavy workflows, and cut the Python container image by 64% during the platform upgrade.
- Test coverage from 45% to between 80% and 95%: automated tests and SonarCloud across the platform's Python components, with zero production defects. Predictor auto-refresh and self-healing model loading kept customer SLAs.
- GenAI prototype, pitched to R&D leadership: a Q&A prototype over internal documentation on LangChain and Pinecone, presented at an internal innovation event in 2023.
- Helping engineers get up to speed: weekly Apache Spark sessions, first-round ML interviews, mentoring new hires and interns, and the prediction workflow guide and Python best practices that cut invalid support tickets.
Three problems from the early years
- Moving time series flows from Pandas to PySpark without losing speed
- Our Pandas flows ran out of memory when a container could not hold a whole file, and time series data cannot be chunked. After the move to Spark, the seasonality flow took 300 seconds a run because it kept spilling to disk. Stack Overflow had nothing useful, so I worked from an early 2010s article on tuning Spark's garbage collector and JVM runtime, profiled and tested the settings, and brought the run to 30 to 45 seconds, close to the 20 seconds of non-distributed Pandas.
- A notebook platform data scientists could rerun
- Data scientists ran independent, untracked notebooks by hand. Taking the idea from Netflix's scheduled notebooks, I built a similar platform for artifact tracking. Notebooks became reusable, automated and reproducible, could run across many users, and later made the move to Modin easier.
- Designing most of the in-house ML platform as microservices
- Data extraction, data preparation, model building and the rest were built component by component. The hard part was switching between flows at runtime from configuration while staying fast. A singleton was one of the patterns that did it. I broke the requirements into tasks and subtasks and built them one at a time, before GPT-3 existed to help.
Aug 2020 to Jan 2021
Software Engineer Intern, ML
Manhattan Associates
- DVC pilot for ML artifact tracking: the findings shaped roadmap decisions on reproducibility and model governance.
- Service middleware: authentication and service discovery components on Netflix Eureka that made internal services more reliable.
- FastAPI research: documented how it could serve online model maintenance, feeding the real-time ML architecture plan.
Manhattan Associates
Recognition
- Jan 2025
Top Talent
Selected for the accelerated growth track.
- Dec 2023
Star of the Quarter
Performance optimization and ML workflow reliability.
- Dec 2022
Above and Beyond
64% smaller Python container images.
- Apr 2022
Star of the Quarter
15x faster critical ML processing.
- Jul 2021
Rising Star
Early engineering contributions.
- 2023
Google Cloud Professional ML Engineer
Among the earliest cohorts at Manhattan to hold the credential.
Get in touch
© 2026 Sagar Devaraju