Python backend engineer with strong AI/ML depth — backend-first, AI-capable. My core work is production backend systems (FastAPI/Django, PostgreSQL, Kafka, async, microservices, high-load), with ML/LLM integrated as first-class functionality — RAG, semantic search, predictive analytics, computer vision — rather than as a separate research track. I also teach and research applied ML as a Senior Lecturer, and publish and maintain open-source tooling for AI-agent reliability.
Independent, Apache-2.0-licensed Python libraries addressing reliability gaps in LLM-agent tool-calling and autonomous decision-making — each with a CI pipeline, coverage-validation test suite, and real-world validation against live models/agents, not just synthetic benchmarks.
| Package | What it does | Links |
|---|---|---|
| KVCalib | Turns "how aggressively can I evict my LLM's KV cache" into a calibrated eviction-budget decision backed by a distribution-free statistical guarantee (Conformal Risk Control), instead of a benchmark number borrowed from someone else's traffic. Supports multiple eviction backends, group-conditional coverage, and an online drift monitor that flags when the guarantee stops holding. |   |
| DeferGuard | Turns any confidence signal (logprob, judge score, ensemble vote) into a calibrated act-vs-escalate decision backed by a distribution-free statistical guarantee (Conformal Risk Control). Includes multi-expert routing and group-conditional coverage. |  |
| ConformGuard | Wraps a single agent tool call with a calibrated accept/abstain decision backed by split conformal prediction's finite-sample coverage guarantee, replacing hardcoded confidence thresholds. Supports joint calibration across multiple simultaneous checks. |  |
| ResilientForge | Gives tool-calling agents persistent, cross-run failure memory: recovers from a failure once, resolves repeat occurrences with zero additional model calls, and promotes proven fixes into standing guards that prevent the failure from recurring at all. |  |
📄 Full list of peer-reviewed papers and book chapters: a-bahrami.com/publications