Karthik Palani

A review agent that sits on GitLab merge requests. It parses the unified diff, decides where it needs more context, and goes and gets it — pulling whole files, searching the codebase semantically, and recalling how similar changes were reviewed before. Findings land as inline comments with one-click GitLab suggestion blocks, deduplicated across review iterations so it does not repeat itself when a developer pushes a fix.

• Episodic memory Stores past reviews and developer responses, then retrieves the relevant ones — feedback stays consistent across MRs instead of resetting every run.
• Agentic tool use Claude decides when to fetch a full file, run a codebase search, or query memory — rather than being handed a fixed context window and guessing.
• Thread lifecycle Answers follow-up questions in a discussion thread and auto-resolves it once the issue is genuinely addressed.
• Two run modes Long-running service polling a GitLab group for MRs where the bot is assigned reviewer, or a single-shot CLI call for pipeline integration.
• Built to stay up SQLite state so restarts never double-review, thread-pool concurrency, Prometheus metrics, FastAPI dashboard, multi-arch Docker images.
• Swappable brain Provider factory behind the reviewer — the model is an env var, not a rewrite.

QA writes a test case in plain English in TestRail. TestFlow reads it, works out what the steps mean, drives a real Chromium browser through them with Playwright, validates the outcome, captures screenshots, and writes pass/fail back to TestRail. Nobody writes a selector. Nobody sits and watches it run.

• Vector memory Remembers how a step resolved last time and replays it — cutting model calls by roughly 70–80% on repeat runs, which is what makes nightly execution affordable.
• Degrades gracefully With no model key configured it falls back to rule-based parsing (regex plus HTML inspection) instead of failing the run.
• Beyond the browser Targets web HMI, PLC endpoints and GitLab — the same case format drives industrial and software surfaces.
• Evidence trail Full execution history, screenshots and timing metrics per step, streamed live over WebSocket while it runs.

The argument in Data residency picks the topology, taken to its limit and actually built. Cadence records a meeting from the microphone and produces a speaker-labeled transcript and a structured Minutes of Meeting — decisions, action items, risks — with no cloud STT, no cloud LLM and no telemetry. Everything runs on the machine. Writing it meant working below the API line: quantised GGUF weights, RAM budgets, native binaries orchestrated from Rust. That is the part I wanted to learn, and it is what makes me trust the self-hosted half of an architecture diagram.

• Fully local pipeline whisper.cpp and sherpa-onnx for speech-to-text, llama.cpp for the minutes — all local GGUF models, nothing sent anywhere.
• Code-switch aware Handles multilingual input, including English, Tamil and code-switched speech, not just single-language dictation.
• Anonymous diarization Labels distinct speakers in the transcript without identifying who they are.
• Grounded minutes Decisions and action items are generated with timestamp references back into the transcript, not free-floating summary text.
• Model integrity + sizing SHA-256 verification on every model it manages, with RAM-aware selection so it runs on the hardware it's actually installed on.