Ali Naqvi

Engineers were checking Grafana dashboards and filling out Jira test records by hand. I connected Terraform-managed alerts to an authenticated Flask webhook that generates a PDF and attaches it to the test execution through the Jira API.

I also added a reporting mode that uses the alert payload directly, so it can run without a Grafana rendering token. I verified the full flow with a real Jira attachment and added a dry-run mode for local testing.

I built fleet, vehicle, and test-level views in Grafana across three vehicles and five ECU platforms. The queries show PASS, FAIL, IDLE, or N/A, including vehicles that have stopped reporting.

One misleading success-rate panel came down to mismatched units: one counter measured uploaded parts and another measured files. I replaced the percentage with paired counts.

I implemented a Python validation framework for Fluent Bit configurations and integrated merge-blocking checks into GitHub Actions across 5+ ECU platforms.

I built plugin-local YAML contracts and a bundler that combines them into a versioned artifact. The validator runs 11+ checks for parameter types, required fields, bounds, dependencies, platform support, and version compatibility.

Platform labels did not map reliably to library versions, so I added support for declaring the minimum Fluent Bit version directly in a configuration. Older files use a compatibility fallback with a warning.

I aligned Python and C++ cache hashes using SHA-256 and used cache telemetry to route requests to workers with matching prompt prefixes. I then implemented a concurrent radix tree for prefix lookups, with Google Test and Bazel tests for insertion, lookup, and concurrency.

Compared with the previous hash-map design, benchmarks showed 35% lower mean time to first token and 30% higher throughput.

I also fixed a routing bug caused by mismatched instance IDs and endpoints, and a contention issue by sorting thread-local copies of worker statistics.

I integrated separate prefill and decode placement, using cache affinity for prefill and available KV memory for decode.

When requests stalled near the end of a workload, I traced the problem to control-channel conflicts, stale requests, and a decoder waiting for a placement that would never arrive. I separated the channels, corrected buffer accounting, and fixed the final-request handling in the single-producer setup.

TTL eviction, backpressure, and PyTorch allocator tuning helped keep stale tensors from filling GPU memory.

I built a LangGraph and LangChain framework with shared rate limits, token quotas, execution logs, and checkpoint recovery. It supports sequential, parallel, and stage-synchronized execution.

I tested 10+ concurrent agents under simulated API quotas. Parallel runs were up to 3× faster than the sequential baseline.

I led the firmware design for power distribution, using FreeRTOS tasks for sensor polling, fault detection, and CAN communication.

I also wrote an I2C driver for the MCP3221 ADC to monitor voltage and current on the rover’s electronics.