Skip to content
View burntcookiedough's full-sized avatar
🎯
Focusing
🎯
Focusing
9 followers · 17 following

Block or report burntcookiedough

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.

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
burntcookiedough/README.md

LOCAL_INFERENCE / AGENT_HARNESS / FAILURE_RECOVERY

Anshul Panigrahi

AI/ML Engineer building edge AI systems and tool-using agents



Current build loop


AI/ML internships | agent systems | edge inference | 1 patents published

Important

Actively looking for AI/ML internship roles where edge inference, tool-using agents, and industrial data matter more than demo-only chatbots.

patents
1 published		 scale
5,000 machines		 edge
<100ms inference		 agents
recovery pipelines		 
[profile]
name = "Anshul Panigrahi"
role = "AI/ML Engineering Intern"
location = "VIT Vellore, Tamil Nadu, India"
current_build = "Hermes Agent + custom tool harness"

[operating_rules]
cloud = "optional"
failure = "recover, verify, hand off"
target = "edge AI, industrial ML, embedded inference"

Repository Map

Track Repositories
Flagship AI/ML Predictive Maintenance (private/on request) · Mode Discovery · Stella · Veri-Dose
Agents + Systems Cognitive Load Scheduler · Aether Dashboard
Edge + IoT Vision Air Sim · Air Safety Assistant · Smart Energy
Vision + Security Face Privacy Filter · Sobel CUDA · QR Security · SecureTorrent

Case Files

Real-Time Predictive Maintenance
Patent application filed | Feb 2026

Kafka, Spark, Cassandra, PyTorch, and graph analytics across 5,000 simulated factory machines to predict cascade failures before downtime.

98% accuracy F1 0.96 AUC 0.99 <12ms latency

Repository available on request.

Physical-Bounded Multimodal Mode Discovery

Unsupervised fault discovery on NASA CMAPSS and CWRU. HDBSCAN proposes regimes; physics constraints reject invalid vibration and thermodynamic behavior.

zero labels 7 CMAPSS regimes 8 CWRU regimes

Repository

Stella

On-device health assistant with Mistral 7B via Ollama. Wearable anomalies become local context, keeping personal health queries off the cloud.

29 metrics 33 users 100% local 0 cloud calls

Repository

Veri-Dose
Patent published | Mar 2026 | IN202641027860 A1

Smart medication dispenser using quantized MobileNetV2 on Raspberry Pi 4. Low-confidence predictions route to a human fallback.

<100ms inference offline Raspberry Pi 4

Repository

Agent Harness

I am not building chatbots. I am building systems where models plan, call tools, recover from failure, and hand off when another model or tool is a better fit.

open agent harness

The model is not the product. The harness is: context routing, tool contracts, verification, recovery policy, and handoff logic.

flowchart LR
    A["input"] --> B["context router"]
    B --> C["planner model"]
    C --> D["tool call"]
    D --> E{"verified?"}
    E -->|yes| F["ship action"]
    E -->|no| G["recover"]
    G --> H{"specialist needed?"}
    H -->|yes| I["sub-agent"]
    H -->|no| C
    I --> F
Loading

Current stack: N8N, Claude Code, OpenAI Codex, Cursor Agent SDK, OpenClaw, Hermes, and custom tool pipelines when frameworks get in the way.

Stack


Python PyTorch TensorFlow FastAPI Apache Kafka Raspberry Pi

GitHub Signals



Contribution Snake

contribution snake

Certifications

Certification Issuer Date
Getting Started with Deep Learning NVIDIA DLI Mar 2026
OCI 2025 AI Foundations Associate Oracle Mar 2026
Software Engineer Intern Certificate HackerRank Feb 2026
Intro to Machine Learning Kaggle Feb 2026

Competitions

If your model needs the cloud to run, we have different philosophies.


If your agent cannot recover from a bad tool call, it is not an agent. It is a very slow API.



Profile Views

Pinned Loading

  1. Stella Stella Public

    Local-first AI health assistant using wearable metrics, anomaly detection, and an Ollama LLM backend.

    Python 1

  2. Veridose Veridose Public

    Offline smart medication dispenser with Raspberry Pi, quantized MobileNetV2 pill classification, and calibrated fallback.

    Python 1 1

  3. AEGIS-ZONE AEGIS-ZONE Public

    Full-stack embedded safety and security platform with backend, frontend, and embedded modules.

    TypeScript 1

  4. Physical-Bounded-Multimodal-Mode-Discovery Physical-Bounded-Multimodal-Mode-Discovery Public

    Physics-constrained unsupervised fault-mode discovery for CMAPSS and CWRU using HDBSCAN and multimodal features.

    Python 1

  5. SecureTorrent SecureTorrent Public

    Malware-aware torrent security prototype with YARA rules, hybrid ML detection, qBittorrent integration, and API tests.

    Jupyter Notebook 3

  6. Cognitive-Load-Aware-Distributed-Task-Scheduler Cognitive-Load-Aware-Distributed-Task-Scheduler Public

    Distributed task scheduler prototype that uses cognitive-load signals for adaptive workload assignment.

    Python