// ascii.mesh
render_cycle: 17.5s
[Yaseen Khalil]|Computational Modeler & ML Systems Architect
> _
$ cat ./tech_tree
Tech Tree
tree
└── active-stack/
├── modeling/
│ ├── python
│ ├── pytorch
│ ├── numpy
│ ├── gnn
│ └── stp-algebra
├── backend/
│ ├── go
│ ├── fastapi
│ ├── postgres
│ └── kdb-q
├── frontend/
│ ├── typescript
│ ├── react
│ ├── tailwind
│ └── nextjs
├── research/
│ ├── r
│ ├── depmap
│ └── systems-biology
└── infra/
├── docker
├── vercel
└── supabase
READY| SCORE: 0| HP: ■■■
$ ls ./projects --verbose
Systems Architecture
- >Created FlyGPlot (fly-g-plot.vercel.app), a Drosophila transcription-factor explorer built around the lab's single-cell workflow: expression data flows through transcription-factor → co-expression module → functional enrichment stages, with a ledger-style interface and live integrations to FlyBase, Ensembl, STRING-DB, g:Profiler, Enrichr, and NCBI GEO so each stage hands off cleanly to the next.
- >Researching how to improve split-GAL4 driver design: current hemidriver selection relies on binary overlap heuristics (Jaccard/specificity indices over thresholded expression patterns), which throw away the strength and shape of gene co-expression. I am designing a KSG k-nearest-neighbor mutual-information pipeline that ranks candidate AD/DBD pairs on continuous single-cell expression from the developing Drosophila optic lobe, capturing the threshold and saturation effects that standard correlation metrics miss.
- >The validation plan is to first confirm KSG-derived rankings recover known, previously validated split-GAL4 combinations from the literature, then benchmark against Spearman-correlation and Jaccard baselines to quantify whether the added complexity measurably improves true-positive recovery — with the goal of proposing novel driver pairs for visual neuron types that currently lack genetic access tools.
Machine LearningFlyGPlotscRNA-seqKSG-MISplit-GAL4
Research Ongoingview_flygplot
- >Developed STP-Diff, a framework that makes Boolean gene-regulatory networks differentiable: discrete dynamics are mapped onto a continuous manifold via Semi-Tensor Product (STP) representation with an implicit Vector-Jacobian Product (VJP) operator, cutting the exponential cost of explicit STP (O(4^N)) down to O(N·2^N) — a measured 353x speedup at N=20 with no numerical approximation, keeping attractor analysis tractable past the ~15-node wall.
- >Applied Projected Gradient Descent (PGD) adversarial attacks — borrowed from computer vision — to find minimal logical perturbations that force phenotypic transitions, exposing the Rb-E2F axis as the vulnerability bottleneck of a 10-node mammalian cell-cycle model. The prediction held up empirically: stratified CRISPR-Cas9 knockout data from DepMap 25Q3 (1000+ cancer cell lines) showed Rb-loss lines become critically dependent on E2F (p < 10^-6).
- >Built an Epsilon-Critical search that measures whole-network resilience as a single number, establishing ε_critical = 2.8000 for the p53-Mdm2 DNA-damage circuit — the perturbation magnitude where the network's negative feedback stops absorbing attacks and the apoptotic attractor takes over. Published as a preprint: DOI 10.13140/RG.2.2.35274.32965.
PyTorchSTP-DiffSemi-Tensor ProductAdversarial MLSystems Biology
Preprintview_paper
- >Built Talii to replace outdated barber platforms: fragmented booking, no client retention, and no-shows plague the industry, so I co-founded and shipped a unified scheduling, client-management, and discovery platform end-to-end, growing it to a live user base.
- >Proposed the operations system the company runs on today — designed the core workflow from scratch and presented it to multiple shareholders.
- >Engineered the whole ecosystem on a unified TypeScript mono-repo with 110+ modular components, keeping feature parity across web and native mobile so barbers and clients get the same experience on any device.
- >Designed a normalized PostgreSQL schema with RLS-based security for distributed scheduling, entity locations, and analytics — permissions enforced at the database layer rather than in app code.
TypeScript Mono-repoReact NativePostgreSQLRLS Security
- >Commissioned by La-Unica, a restaurant brand, to design and build their complete digital presence — a full proposed design delivered under commission, covering the customer-facing experience and the operational system that runs the business day to day.
- >Worked across the whole operational stack: the front end for the proposed site plus the systems behind service, ordering, and day-to-day operations, built out as one connected system.
DesignOperationsReactFrontend
CommissionedCommissioned Project
- >Joined the founding team of a venture-backed real estate startup as the machine learning lead — a project with multi-million-dollar VC backing and a pipeline of clients already waiting.
- >Lead all machine learning work across the platform. Specific architecture and model details are under NDA.
Machine LearningReal EstateConfidential
Under NDAdetails under NDA
- >Designed and deployed an end-to-end vehicle health monitoring neural network — a bi-directional LSTM with attention — that catches developing faults in fleet telemetry before they become breakdowns.
- >Iteratively retrained on 30K+ telemetry points every 10 minutes, lifting anomaly detection accuracy from 40-60% to 87-95%.
- >Architected a three-tier detection pipeline — LSTM for sequential insight, DBSCAN for geographic hotspot clustering, and FastAPI microservices on Railway with Supabase feeding an internal dashboard UI.
BiLSTMAttention MechanismsDBSCANFastAPISupabase
- >Built the front end for Virginia Tech's Diggeridoos robotics team, ingesting Arduino telemetry and turning it into live visualizations so the team could see how the machine was actually operating in real time.
- >Led the movement-based visualization work: a 360° sonar-style heading view with a position dot that made facing direction and turning legible at a glance — operators could see where the machine was pointed instead of reading raw sensor values.
- >Worked on a team using React and modern front-end tooling to convert raw sensor streams into an intuitive operator dashboard.
ReactArduinoData VisualizationRobotics
Academic Project
$ git log --blog --oneline
Blog
Published writing
blog.log
5 entries |5 mergedHEAD -> main