Skip to content
View sepahead's full-sized avatar
🐳
building something great.
🐳
building something great.

Block or report sepahead

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
sepahead/README.md

Sepehr Mahmoudian, Senior AI Engineer in Berlin building domain-specific AI agents and custom harnesses for auditable research and engineering, with work spanning LLM/VLM evaluation, computational neuroscience, robotics and multimodal 3D perception. Rust and Python; GitHub member since 2014.

Sepehr Mahmoudian on LinkedIn — Senior AI Engineer in Berlin    Sepehr Mahmoudian on Google Scholar — research publications    Sepehr Mahmoudian on Substack — @torusprime, notes and essays    Sepehr Mahmoudian on Hugging Face — @torusprime, datasets    Sepehr Mahmoudian on X (Twitter) — @SepAhead

Curriculum Vitae · English  ·  Lebenslauf · Deutsch

The pulse — GitHub contribution activity

Annual contributions from 2014 to present, with 2014–2022 grouped into a single history bar and a running cumulative total; the highest year so far glows, its in-progress status is explicitly qualified, the year in progress carries a dashed cap, average growth is qualified with its base year, the next year sits as an empty dashed placeholder and the following year as an unstarted future runway slot, with the visual phase motif resolving between them into 20 rays of light streaming from above into the future.

Weekday contribution share: a donut of each weekday's percentage of the last ~16 months (500 days) of contributions; the glowing slice marks the peak day.

Selected work — computational neuroscience, robotics, 3D and scientific software projects

engram: Engram experiments use biophysical and functional neural models. An AI agent and custom harness track model sources, simulation inputs, and results. Its optional NCP controller matched direct NEST readouts in a CREBAIN loop. The public repository remains a placeholder pending source publication. Private Paper2Brain source does not claim general paper reproduction or validated scientific results. NCP: NCP defines typed messages for neural controllers, simulations, sensor monitors, and experiment capture. Applications select local peers through private process channels. Installed sensor and NEST loops passed bounded checks. Broader qualification and final v1 release gates remain open. Haldir retains its separate pinned v0.8 interface. Languages: Rust, TypeScript, Python, C, C++.
prisoma: Prisoma develops experiments for action-conditioned world models. Its Agent Bridge records commands and original NCP exchanges. The native pressure study completed 112 episodes with forecasts and restored outcome labels. Forecast improvement missed the declared useful threshold. LeWM inference and search passed one frozen CPU/MPS case. General model quality and learned action benefit remain open. PID is an optional diagnostic. Languages: Rust, Python. crebain: CREBAIN provides standalone drone simulation with RGB images, microphone pressure, thermal radiance, and optional NCP interfaces. Native checkpoint controls passed with 3 and 16 endpoints. City tests exercised 256 drones through NCP. Tracking qualification remains open. Physical realism and learned-policy quality are not established. Languages: Rust, TypeScript.
melkor: Melkor provides deterministic 3D Gaussian Splatting (3DGS) conversion across PLY, SPZ and glTF. Its inspection reports surface field provenance, bounds and numeric hazards without modifying source assets. The current release candidate is source-only; no production binary is currently supported. Languages: C++, Python, JavaScript. galadriel: Galadriel is an optional Rust monitor for possible sensor tampering and persistent measurement anomalies. It checks prediction residuals, changes over time, and agreement between sensor modalities. Faults, model errors, or tampering can cause an alert. An alert alone does not identify the cause. Its NCP adapter records assessments without command authority. The current single-Visual integration test correctly reports insufficient evidence; multisensor tampering qualification and field validation remain open. Languages: Rust.
pid-rs: pid-rs is a safe-Rust library for shared-exclusions Partial Information Decomposition and mutual-information estimation: categorical SxPID, KSG MI and default-off experimental continuous shared-exclusions/PID surfaces. v0.9.0 is a GitHub-only source-review prerelease. Languages: Rust, Python. haldir: Haldir Gate researches mission authorization, signed controller intents, admission, deterministic policy, and decision receipts. Its 0.9 review remains NO_GO. It is outside the NCP local v1 simulation profile; gated requests must be rejected before endpoint preparation. Local NCP digests are not Haldir signatures. No production readiness or airworthiness is claimed. Languages: Rust.
manwe: Manwe is an alpha airspace-perception research and validation workbench spanning vision, audio, multi-camera geometry and multi-target tracking, with a Python numerical/training package and Rust/Candle inference benchmarks. It targets candidate outputs for systems such as Crebain but currently ships no drop-in Crebain adapter. Languages: Python, Rust. cortexel: Cortexel is an unreleased TypeScript library and CLI for neural-simulation figures: it validates and canonicalizes strict declarative JSON requests with fail-closed provenance, then returns deterministic SVG plus a complete exact-value table. main is 0.10.0-dev.0; v0.9.0 is the last tagged preview. Languages: TypeScript.
relief-atlas: relief-atlas is a Python generation pipeline and 10,079-item manifest and prompt catalog for disaster-relief, humanitarian-aid and civil-protection meshes. The repository currently contains 125 GLBs, not a complete 10,079-mesh corpus. Languages: Python. cobot-atlas: cobot-atlas is a Python generation pipeline for a public MIT dataset of 2,023 unique glTF 2.0 Binary meshes (2,024 GLB files, 33.5 GB) for robot/cobot simulation, manipulation research, VLA training and benchmarking. Languages: Python.

How the work relates

The map shows protocol interfaces, libraries, tools, and assets. Applications select the components they need.

NCP connects Engram, CREBAIN, Prisoma, and Galadriel through local interfaces. Haldir retains its pinned v0.8 interface. A long dashed open arrow joins CREBAIN sensors to Prisoma experiments. The native pressure study completed with a null or inconclusive forecast result. Short dashed arrows show libraries, dotted lines show assets or exports, and moving dashes show perception tools.

Open diagram · zoom and explore · Original SVG: light · dark

Read the connections: solid paired arrows = local NCP interfaces; short dashed arrows = library dependencies; long dashed open arrow = environment and sensor integration, with the sensor path tested; dotted lines = assets or exports; moving dashes on a continuous line = perception tools; dash-dot line with a square end = Haldir's pinned NCP v0.8 interface.

Different jobs: Engram runs neural models. CREBAIN runs a standalone 3D environment and sensor fusion. Galadriel optionally checks sensors for possible tampering. Prisoma organizes experiments and evidence for embodied agents. NCP defines their shared messages.

Tested neural loop: Engram ran a real NEST network through NCP with CREBAIN. The loop ran 8 neural steps and 24 body ticks. All 64 readouts matched direct NEST execution. The test used explicit network and scene inputs. Paper-derived model reproduction remains a separate task.

CREBAIN and Prisoma: Prisoma records commands before CREBAIN executes them. Installed body-only and canonical runs matched all 44 payloads, totaling 4,326,400 bytes. A 192-run timing study missed every 120-Hz deadline. The native pressure study completed 112 episodes with restored labels. Forecast improvement missed the declared useful threshold.

Four-owner local v1 reference. Local sensor and neural loops passed their declared checks. Broader qualification and final v1 release gates remain open.

The tested Engram implementation is private Paper2Brain source. The public Engram repository is a placeholder, not an executable release. Read the NCP release status.

Components, sensors, and operating limits

CREBAIN runs standalone. Applications can add Engram neural models, Prisoma recording, or Galadriel monitoring. One endpoint can expose several identified sensors.

Each camera and microphone retains its identity and timing. Prisoma defines features, source groups, targets, and statistical assumptions. PID-rs estimates the declared quantities. The initial PID study selects two through four source variables; NCP imposes no such limit.

Melkor converts splat assets. The atlas projects provide mesh assets. These are intended CREBAIN inputs; the full import path remains unqualified.

The four-owner reference routes CREBAIN observations to Engram and returns its action proposals. Galadriel records diagnostics, and Prisoma captures complete exchanges. Applications can select smaller compositions through the modular interface.

Engram coordinates one local experiment and privately owns neural state. CREBAIN owns body and fusion state. Prisoma captures exact step pairs. Galadriel records actual detector output with explicit abstention.

Each owner exchanges bounded requests and results through its own private process channel. NCP defines the shared contract.

Galadriel's current detector requires two sensor modalities. Two RGB cameras supply one Visual modality. Missing evidence produces abstention, never a nominal report.

Supported boundary: This candidate targets local Darwin simulation. Haldir gating, remote endpoints, physical actuation, and real-time guarantees are excluded. Gated requests must be rejected before endpoint preparation. Capture and monitor results grant no command authority.

Engram coordinates one local experiment and privately owns neural state. CREBAIN owns body and fusion state. Prisoma captures exact step pairs. Galadriel records actual detector output with explicit abstention. This candidate targets local Darwin simulation. Haldir gating, remote endpoints, physical actuation, and real-time guarantees are excluded. Gated requests must be rejected before endpoint preparation. Capture and monitor results grant no command authority.

Open diagram · zoom and explore · Original SVG: light · dark

More repositories — public research code and tools


More public repositories; a detailed link list follows.

↗  brojapid-activationfunctions · mahmoudian-2020-rescience · nest-simulator · relief-atlas · silmaril-vision-studio

The toolbox — languages, frameworks and infrastructure


AI / ML stack: Python, PyTorch, NumPy, Pandas, SciPy, Pydantic, scikit-learn, Jupyter
Python    PyTorch    NumPy    Pandas    SciPy    Pydantic    scikit-learn    Jupyter
Backend & Systems stack: Rust, C, C++, FastAPI, Drizzle ORM, PostgreSQL, gRPC, Zenoh
Rust    C    C++    FastAPI    Drizzle ORM    PostgreSQL    gRPC    Zenoh
Cloud & DevOps stack: Cloudflare, Google Cloud, AWS, Docker, Kubernetes, Terraform, GitHub Actions, Linux
Cloudflare    Google Cloud    AWS    Docker    Kubernetes    Terraform    GitHub Actions    Linux
Frontend & Web stack: JavaScript, TypeScript, React, Vite, TanStack, Tailwind CSS, Vitest
JavaScript    TypeScript    React    Vite    TanStack    Tailwind CSS    Vitest

Agentic engineering — the AI-agent development stack

Agentic stack manifest: Ghostty (terminal, GPU-native); herdr (multiplexer, agent herd); OMP (lead harness, turbocharged Pi with batteries included); Devin (harness, long-horizon); Zed (editor, collaborative IDE).

Elsewhere — contact channels

Open channel: reach me at sepmhn@gmail.com; always open to interesting problems.

Pinned Loading

  1. crebain crebain Public

    Drone simulation with RGB, acoustic, and thermal sensors, native checkpoints, and optional NCP interfaces. Research prototype.

    Rust 22

  2. engram engram Public

    Engram Neural Modeling Labs: NEST and NCP experiments. Public source is pending; this repository is a placeholder.

    15

  3. NCP NCP Public

    Typed local protocols for neural simulators, sensors, and experiment capture. NCP v1 qualification is in progress.

    Python 16

  4. melkor melkor Public

    C++17 toolkit for 3D Gaussian splats: deterministic conversion, inspection, geometry-based scene completion and viewing across PLY, SPZ and glTF, with CPU/Metal and optional CUDA backends. v2 harde…

    C++ 13 1

  5. pid-rs pid-rs Public

    Safe Rust estimator core for categorical shared-exclusions PID (SxPID), KSG mutual information, and default-off experimental continuous PID. Optional Python bindings. v0.9.0 source-review prerelease.

    Python 15

  6. cobot-atlas cobot-atlas Public

    Python generation pipeline behind a public MIT dataset of 2,023 unique glTF 2.0 Binary meshes (33.5 GB) for robot/cobot simulation, manipulation research and VLA training. Published on Hugging Face…

    Python 7