Context
Tempo's hundreds of clients were siloed across separate databases and shards, which meant patterns were invisible and a real machine-learning solution was out of reach. Meanwhile, the core product problem — getting engineers to log accurate timesheets — remained a manual, unreliable chore.
What I Did
- Personally defined the evaluation criteria and success metrics used to determine which models and systems were actually working, rather than inheriting someone else's definition of success.
- Directed development of an API gateway and data pipeline feeding a Snowflake data warehouse, unifying core database updates and plugin behavioral events into a single location for the first time.
- Directed a systematic evaluation of NLP and ML approaches — including LSTM, LDA, XGBoost, SVM, Naive Bayes, and embedding-based cosine similarity — to detect missing timesheet entries from users' Jira, calendar, and code-editing activity.
- Expanded ML-based capabilities into existing web and mobile applications, and built entirely new Slack and Microsoft Teams interfaces so the capability lived where the team already worked.
- Coordinated executive, investor, product, design, and operations alignment on a 12-month roadmap and capability prioritization.
Outcome
- Over 70% suggestion-acceptance accuracy on automated timesheet detection
- A functioning ML center of excellence stood up from a cold start, with a team staffed around it
- “Kickdrum's expertise was invaluable in bringing machine learning to our product.” — Mark Lorion, CEO, Tempo Software
Capabilities Demonstrated
ML Evaluation & Success Metrics
Natural Language Processing
Data Pipeline Architecture
Cross-Platform Integration
