Al Harith Abu Mahfuz

• Designed and shipped the MVPs core GenAI chat feature on Google Vertex AI using Gemini Flash
2.5 and the Google GenAI SDK, tested by 72 internal employees during the closed pilot.
• Built a tool-calling architecture with three modular triggers: prompt-injection protection,
conversational mood tracking, and a critical-state escalation flag that alerted clinical admins.
• Implemented chat-specific session memory tuned to roughly 20-exchange conversations, with
handoff logic to prevent termination at sensitive moments and groundwork for cross-session
context.
• Built an in-product feedback loop (Helpful / Not what I asked for plus free-text follow-up) and
fed user signals back into prompt refinement.
• Navigated data-residency constraints (Googles KSA region lacked LLM support at the time,
requiring US-region deployment) and documented the compliance trade-offs for stakeholders.
• Collaborated with backend engineers on PostgreSQL schema design and with React Native
developers on integrating AI output signals into the onboarding flow.

• Built a Retrieval-Augmented Generation proof-of-concept over internal tasks tickets using
Python, Flask, LlamaIndex, and Chroma DB, deployed on on-premises infrastructure with a 32
vCPU allocation.
• Designed a selective-indexing strategy: only unresolved tickets were embedded, with re-fetch
and upsert on ticket changes, to stay within VPC storage limits without losing recency.
• Achieved millisecond vector retrieval (~0.2s per query) by careful index sizing and tuning.
• Wrote a static Python pre-filter that resolved intent ambiguities before LLM inference,
improving relevance of retrieved context.
• Validated retrieval quality through manual cross-checks across the web portal, CLI, and RAG
output, reviewed by four team members and the director during PoC presentations.
• Integrated the Python inference layer with existing PHP and MySQL backend services and
executed supporting database migrations.

• Researched transformer architecture, semantic analysis, weight distribution, and prompt
engineering patterns, documented in a structured internship report covering applied LLM use
cases.
• Lifted a garbage-classification computer vision model from approximately 80% to 99.6%
accuracy by restructuring training weights, adjusting epoch distribution, and experimenting
with alternative probability functions.
• Assisted in dataset curation (training and testing image sets) and advised peers on ML
experimentation methodology and prompt-engineering approaches.