This is a submission for the Gemma 4 Challenge:
Build with Gemma 4
What I Built
MySeniorDev is a local-first AI code reviewer powered
by Gemma 4 E2B. You paste any file from your project
and get an honest, specific code review — the kind a
senior developer gives in a real pull request.
Three review modes:
- Security — finds hardcoded secrets, missing validation, exposed API keys, auth holes
- Architecture — flags god files, mixed concerns, patterns that won't scale
- General — full review across security, structure, and code quality
After the review you can ask follow-up questions.
Gemma 4 holds your full file in context for the
entire conversation.
Built for self-taught developers who have never had
a mentor or senior developer to learn from. In Nigeria,
India, Indonesia — anywhere bootcamps are expensive
and code reviews are not accessible.
Demo
Code
The core of the project lives in two places:
The system prompt — backend/services/prompt.js
This is where the real intelligence is. Gemma 4 is
instructed to behave like a senior developer doing
a real pull request review — direct, specific, always
referencing the actual code, never giving generic advice.
The Ollama integration — backend/services/ollama.js
Sends code to Gemma 4 E2B running locally via Ollama
at http://localhost:11434. Falls back to Google AI
Studio automatically if Ollama is unavailable.
How I Used Gemma 4
Model chosen: Gemma 4 E2B via Ollama
I chose E2B for one reason: it runs on hardware most
developers in the world actually own. 8GB RAM. No GPU.
Consumer laptop. That is the target user's machine and
the model had to match.
Gemma 4 is doing real work at three layers:
1. Code Understanding
The model reads entire files pasted by the user and
reasons about them structurally — not just syntax,
but design patterns, security implications, and
architectural concerns. This requires the kind of
contextual reasoning Gemma 4 handles well.
2. Plain Language Translation
The system prompt instructs Gemma 4 to explain every
technical term it uses in the same sentence. A beginner
reading "SQL injection vulnerability" immediately gets
"that means an attacker can type code into your form
and your database will execute it." That translation
layer is entirely prompt-driven — Gemma 4 does it
consistently without being explicitly told each time.
3. Conversational Memory
Gemma 4's context window holds the full file across
the follow-up conversation. When a user asks "how do
I fix the auth issue you flagged?" the model still
knows their exact code. It answers specifically,
not generically.
Why local matters for this use case:
Every other AI code review tool sends your code to
a server. For a beginner pasting real production files
with real API keys and real secrets, that is a genuine
security risk most beginners don't know exists.
MySeniorDev runs Gemma 4 entirely on the user's machine.
Code goes from their editor to their RAM and back.
Nothing else.
That's not just a technical choice. That's the reason
this tool should exist.
Built by @VEND321
YouTube: darkshard
Location: Nigeria
Hardware: 8GB RAM, no GPU, consumer laptop
Model: Gemma 4 E2B via Ollama
Top comments (0)