AI / Software Engineer · Researcher

Hi, I'm Hashem Jaber.

AI/Software Engineer at Akamai Technologies and M.S. in Artificial Intelligence at San José State University. My thesis was defended and approved in November 2026. I work on vision-language-action models, flow-matching and diffusion planners, and multi-modal perception for autonomous driving.

Portrait of Hashem Jaber
  • NowAI/SWE Engineer, Akamai Technologies
  • ThesisDefended & approved, Nov 2026
  • HonorBertha Kalm Scholarship nominee, 2025–2026

About me

I'm a software and deep learning engineer who likes problems where research and production meet. At Akamai I build agentic GenAI systems on the A2A and MCP frameworks, and in my research I study how language reasoning should steer generative planners for autonomous driving.

I finished my M.S. in Artificial Intelligence at San José State University, where I defended my thesis on language-guided flow-matching driving planners under Dr. KaiKai Liu. Before that I earned my B.S. in Computer Science (Software Engineering) at Sacramento State, magna cum laude, and led deep learning work on 6G mmWave beam alignment.

Much of my work starts from a social problem: earthquake early warning, neurodivergence diagnosis, or finding lost items. I'm happiest when a model leaves the notebook and helps someone.

Experience

  1. Dec 2025 – Present

    AI / Software Engineer (L2) · Akamai Technologies

    • Architected a production GenAI platform on the A2A and MCP frameworks with a React management UI; extended the RAG pipeline with an image-aware crawler, driving a 280% engagement increase and 0.81 F1 retrieval across heterogeneous document corpora.
    • Fine-tuned LLMs and VLMs with LoRA and built a KL-divergence evaluation lab measuring how prompt variations shift judge scores, latency and quality against baseline traces.
    • Shipped a Judge Framework with adaptive sampling for automated quality control over LangFuse traces, plus a Debugger Agent for root-cause analysis across Kubernetes logs.
  2. Jun – Aug 2025

    Advanced Data Science Intern · ZS Associates

    • Architected a deep-learning recommender serving 9M+ weekly users (embeddings, collaborative filtering, matrix factorization), with PySpark feature engineering and XGBoost classifiers over behavioral segments.
    • Built client-facing LLM reasoning and debugging tools in the MAX.AI platform so engineers could trace model outputs across production workflows.
  3. Feb – Oct 2025

    Research Lead · Neural Divergence Diagnosis (3-company initiative)

    • Led a multimodal diagnostic system combining speech-to-text, LLMs, VLMs, agentic AI, text-to-speech and human-robot interaction; correlated real-time affective state with eye-movement signals and quantified ADHD severity via a time-normalized topic-switching rate.
  4. Feb – Aug 2024

    Deep Learning Technical Lead (RA) · Sacramento State, EEE Lab

    • Advanced 6G mmWave beam-alignment research with CNNs and Transformers using custom losses and sampling strategies, cutting training-data requirements by 73%; used GANs to address signal-data scarcity. Worked under Dr. Eltayeb.
  5. Sacramento State

    Board Director, Engineering & Computer Science · Associated Students, Inc.

    • Served on the board of a $12M student non-profit and launched the first AI hackathon in Sacramento State's history.

Research

My research sits at the intersection of language, generative models and decision-making for embodied agents.

Research areas

Vision-language-action models & planning

How language reasoning from a VLM should condition continuous trajectory generation in flow-matching and diffusion planners for autonomous driving.

VLA · flow matching · diffusion policies · nuPlan · NAVSIM

Generative world models & multi-modal perception

Cross-modal LiDAR ↔ camera generation, latent alignment and BEV perception that stay robust when sensors drop out.

latent diffusion · VAEs · InfoNCE · BEV · Argoverse 2

Scene representation for VLMs

Learned, interpretable tokenizations of driving scenes that let language models reason about spatial layouts.

scene tokens · VQA · grounding · BDD100K · KITTI

LLM agents & evaluation

Agentic systems, retrieval and the evaluation methods that tell you whether a change actually helped.

A2A · MCP · RAG · LoRA · LLM-as-judge

AI for social good

Earthquake forecasting from animal-behavior time series, neurodivergence diagnosis and deceptive-review detection.

time series · multimodal · NLP

Deep learning for wireless

Data-efficient beam alignment for indoor 5G → 6G mmWave systems.

CNNs · Transformers · GANs · mmWave

M.S. Thesis · San José State University · Defended & approved Nov 2026

Admitting Language Directives Mid-Solve: Scheduled Command–Reason Guidance for Flow-Matching Driving Planners

Chair: Dr. KaiKai Liu · Committee: Dr. Wencen Wu, Dr. Younghee Park · Department of Computer Engineering

Question. When a vision-language model decides something explicit, such as "stop because of a red light", how should that decision shape the continuous trajectory a generative planner is producing?

Approach. An Eagle2-1B VLM, frozen apart from LoRA adapters, reads the camera frame and scene prompt, and its hidden states join map and ego features as scene tokens for a flow-matching planner. Halfway through the ten-step solve, the same language model starts decoding command–reason directives. A buffer admits the valid ones, and a learned per-channel gate blends attention to the scene with attention to the latest directive. One variant, T1, also feeds back how far the directives have moved the plan. The planning loss never updates the language model.

Architecture diagram with a solve schedule. Ego history, an HD map raster, a front camera frame and a scene prompt are encoded into scene tokens; the Eagle2-1B VLM is frozen apart from LoRA adapters. The same language model decodes command–reason directives and an admission buffer keeps the valid ones. In each of four denoiser layers, cross-attention to the scene tokens and cross-attention to the latest directive run in parallel after self-attention, and a learned per-channel gate blends them before the feed-forward layer and the Euler update. Schedule: steps at flow time 0 to 0.4 use scene tokens only; directives are admitted from 0.5 to 0.9; T1 measures the trajectory difference after the 0.6 step and injects it into steps 0.7 to 0.9. The output is a planned trajectory of 80 waypoints over 8 seconds.
The planner at inference, simplified from Figures 1 and 2 of the thesis. The two cross-attention branches run in parallel and a learned gate blends them. The first five Euler steps use scene tokens only, directives are admitted from t = 0.5, and T1 injects the measured trajectory difference into the last three steps.

Evaluation. Four models trained for 100 epochs each on 40,000 nuPlan scenes, with a session-disjoint split (3,400 validation / 2,000 held-out test scenes), validation-only selection across 400 checkpoints, test conditions locked before test inference, and paired log-bootstrap intervals on NAVSIM PDMS.

Finding. Directives measurably change the generated trajectories but did not improve planning: the scene-only baseline scored 81.7 PDMS versus 76.1–77.0 for directive models, with losses concentrated in collision and drivable-area checks and roughly 6× inference latency. Eight-second imitation error alone did not reveal this gap.

40k
nuPlan training scenes
400
checkpoints, validation-only selection
81.7
PDMS, scene-only baseline
~6×
latency cost of directives

The full thesis will be available through SJSU ScholarWorks once it is published.

See it drive

Replays of plans from the thesis models on validation scenes. Pick a scene, then switch models to compare the same moment: the orange car is the model's plan, the green car is the human driver, and the right-hand panel lists the directives the model admitted.

Scene
Model

Validation scenes from the nuPlan dataset, replayed at quarter speed. Surrounding traffic is logged and doesn't react to the model. Each clip is one saved plan from an epoch-60 checkpoint, so per-scene scores are illustrations; the thesis's held-out test results are the evidence. Camera imagery and driving logs: nuPlan (Motional).

Other research

Multi-Modal Generative World Models for Sensor Imputation Unpublished manuscript · 2025–2026

A bidirectional LiDAR ↔ camera generative framework that imputes missing modalities under sensor dropout: a multi-layer range-image VAE, InfoNCE latent alignment over frozen Depth Anything V2 / YOLOv8 / ViT backbones, and latent diffusion for cross-modal generation. A Temporal BEV Transformer over four frames of history improved downstream BEV occupancy IoU by 15% on Argoverse 2.

"Language of Driving": Two-Tower VLM with Learned Scene Tokenization Unpublished manuscript · 2025–2026

A count-aware scene-token codebook, a BPE analogue for spatial layouts, that clusters multi-object descriptors into learned centroids decodable back to pixel-space boxes, with an Aligner network correcting quantization error. Validated on VQA, object localization and latent-diffusion scene reconstruction on BDD100K and KITTI.

EarthquakeNet: Neural Networks for Earthquake Prediction Provost's Award for Research Excellence · 2024

Forecasting seismic events minutes to hours ahead from animal-behavior time-series data. 1st place, graduate division (Engineering/Science/Math), Sacramento State Research Symposium.

Deep Learning for Indoor 5G → 6G mmWave Sacramento State EEE Lab · 2024

Unpublished manuscript on data-efficient beam alignment with CNNs, Transformers and GANs.

LENSE: Deceptive Review Detection Collaborator

Contributed to research on detecting deceptive online reviews.

Paper previews (earlier, unpublished work)

These papers are not published yet. Other work will be shared once approved.

Honors & awards

  • 2025–2026

    Nominee, Bertha Kalm Scholarship

    College of Graduate Studies, San José State University. Nominated Feb 26, 2025.

  • 2024

    Provost's Award for Research Excellence

    Sacramento State, for EarthquakeNet.

  • 2024

    1st Place, Graduate Division

    Engineering/Science/Math competition, Sacramento State Research Symposium.

  • 2024

    Magna cum laude · Dean's Honor List

    B.S. Computer Science, Sacramento State.

Projects

Skills

Embodied AI & autonomous driving

  • VLA models
  • Flow matching
  • Diffusion policies
  • Latent diffusion
  • BEV perception
  • Sensor fusion
  • LiDAR
  • World models
  • nuPlan / NAVSIM
  • Argoverse 2
  • BDD100K
  • KITTI

Deep learning

  • PyTorch
  • TensorFlow
  • CUDA
  • Transformers / ViT
  • CNNs
  • VAEs
  • GANs
  • Contrastive learning
  • LoRA / PEFT
  • RL (PPO, SAC, RLHF)
  • Scikit-learn

LLMs & agentic systems

  • LangChain
  • LangGraph
  • RAG
  • Qdrant
  • LangFuse
  • MCP
  • A2A
  • VLMs
  • NLP
  • Recommender systems

Systems & data

  • Python
  • SQL / MySQL
  • PySpark
  • Pandas
  • NumPy
  • FastAPI
  • Node.js
  • Docker
  • Kubernetes
  • Helm
  • Linux
  • Git

Web & mobile

  • React
  • Next.js
  • React Native
  • TypeScript
  • JavaScript
  • Swift / SwiftUI
  • WebRTC
  • Supabase
  • Leaflet
  • Java
  • C

Education

M.S. Artificial Intelligence

San José State University · Dec 2026

Thesis on language-guided flow-matching driving planners, defended and approved Nov 2026.

B.S. Computer Science, Software Engineering

California State University, Sacramento · May 2024

GPA 3.66 · magna cum laude · Dean's Honor List

Certificates
  • Generative Adversarial Networks (GANs) Specialization · Coursera · Jan 2024
  • Deep Learning Specialization · Coursera · Aug 2023
  • Intermediate iOS · CodePath · May 2023
  • Structuring Machine Learning Projects · Coursera · Feb 2023
  • Neural Networks and Deep Learning · Coursera · Jan 2023
  • Intermediate Cybersecurity · CodePath · Nov 2022
  • Machine Learning Specialization · Coursera · Nov 2022
  • Unsupervised Learning, Recommenders, Reinforcement Learning · Coursera · Nov 2022
  • Advanced Learning Algorithms · Coursera · Oct 2022

Get in touch

I'm open to conversations about research, AI engineering roles and collaborations.

How this site sees you

A small live demo: the chart shows how many seconds you've spent on each section during this visit. It works best on a laptop or desktop.

Seconds spent per section this visit will appear here.

Who visits: people or programs?

Not every visit is a person. This is my best guess at the split, made from how long each visitor stayed, how many sections they reached, what device they used and how they arrived. Read the notes underneath before trusting any of it.

Loading visitor statistics…

How to read these numbers

  • The labels are guesses. No visitor is confirmed as a person or a program. Each one is scored with a short list of rules, shown in the table view, and anything without clear evidence stays "unclear".
  • My database only sees browsers that run JavaScript. Most crawlers never do, so they are missing here entirely. The programs that do show up are ones that open pages in a real browser, such as link scanners.
  • A burst is not proof. The strongest automated signal is the same browser string appearing on several new visitor IDs in one day. A classroom, an office with identical laptops, or me testing with cleared storage would look the same.
  • Short visits are not proof either. A person who opens the page and closes it within a few seconds scores like a program.
  • The two sources can't be joined. MapMyVisitors knows locations but shares no visitor ID with my database, and the two count different things over different periods. Compare them side by side; don't add them up.
  • Location comes from IP addresses. VPNs and corporate networks move people to the wrong city, and a data-centre town points to a program without proving it.
  • The sample is small and mostly old. Most rows are from 2024, recorded by an earlier version of this site. Time-zone, input and section-order signals only exist for visits after November 2026.
What's collected, and why

This site stores a random visitor ID in your browser and records your browser type, language, visit count, how you arrived (for example ?src=linkedin) and time per section in my database, so I can see what's useful and improve the site. To help tell people from programs it also records your screen size, time zone, whether the browser says it is automated, whether you scrolled, clicked, tapped or typed (not what), and the order of sections you viewed. The charts above only use coarse versions of this, never the visitor ID or the full browser string. Separately, a third-party counter (MapMyVisitors) logs approximate location and browser information; that data isn't stored in my database.

Map of site visitors