I am an ML systems engineer, TensorFlow core contributor, and founder building an independent AI lab dedicated to Sovereign Machine Intelligence.

I believe every individual deserves uncompromising privacy without diluting themselves into centralized systems, and every organization must own sovereign AI: models, data, and agent runtimes they govern locally, without leaking proprietary state to cloud monopolies.

Centralized AI stacks are fundamentally brittle and insecure. Real-world autonomy requires two co-designed layers:

• Agents and Harnesses That Do Not Break: Deterministic execution harnesses and runtimes with cryptographic event hashing, immutable action ledgers, and semantic checkpoints that guarantee zero side-effect duplication and verifiable recovery across multi-hour executions.
• Task-Specific Sovereign Models: Lean, domain-specialized architectures engineered specifically for the job, running privately offline or on-premises with zero external data egress.

Building the engineering stack for production sovereign autonomy.

Building autonomous agent execution harnesses and runtimes that do not break, crash, or duplicate destructive side-effects during long-horizon workflows.

• CONTINUUM · [Live Demo] · [PyPI] · [Docker GHCR]
  Production-grade infrastructure for verifiable semantic recovery of long-running autonomous AI agents.
  Apache 2.0 · Python 3.11+ · 28★ · 55 forks · 58 contributors · 4,230+ tests · 14 Phase 6 recovery scenarios

  Solves the multi-hour agent crash failure mode: replaces naive context replay and token-burning hallucinated recoveries with deterministic, verifiable resumption.

  • Zero Side-Effect Duplication: Cryptographically tamper-evident SHA-256 hash-chained event logs paired with an idempotent action ledger physically preventing duplicate mutations (e.g. double billing, duplicate API writes).
  • Semantic Checkpoints and Environment Revalidation: Resumes from compact, versioned state contracts rather than lossy conversation dumps, enforcing staleness propagation across external dependency graphs before execution resumes.
  • Provenance-Aware State and Seven Recovery Modes: Every fact carries its origin; agent-reported progress is marked REQUIRES_REVIEW (never self-certifying) with sealed next-action recovery contracts.
  • Secure Planning Loop and Drift Detection: Two-signal observation verification escalates high-risk branches, with scheduled periodic revalidation catching mid-run environment drift within one cycle.
  • Deny-by-Default MCP Server (13 Tools): Native drop-in integration for Claude Code, Cursor, and Anthropic toolchains with sub-second deterministic rollbacks.
  • Framework Adapters and Wire Sidecar: Native integrations for LangGraph, LangChain, OpenAI Agents SDK, and an HTTP sidecar bridge with OpenAPI wire conformance for non-Python agents.
  • Rigorous Battle Testing: 4,230+ collected tests and 14 Phase 6 recovery scenarios 100% verified inside live Claude Code agent sessions; automated Docker builds on GHCR and interactive web demo.

• SNAGLINE · [Documentation] · [PyPI]
  Zero-dependency, microsecond-scale real-time failure detection and guardrails for LLM agents.
  MIT · Python 3.10+ · 13★ · 9 forks · 11 contributors · Zero dependencies · 1.70 - 2.43 μs median latency

  Eliminates the silent budget-burn problem: catches infinite loops, tool thrashing, latency spikes, and reasoning collapse in real time in <1ms without expensive LLM-as-a-judge calls.

  • Microsecond-Scale Zero Overhead: Runs $O(1)$ amortized statistical detectors using strictly the Python standard library, benchmarked at 1.70 - 2.43 μs median / 27.71 μs p99 over 200,000 synthetic steps on Apple Silicon.
  • Deterministic Multi-Vector Anomaly Trapping: Real-time detection across infinite loops (sliding-window novelty), error cascades, statistical CUSUM latency regressions, semantic goal drift, tool meltdown/entropy collapse, and token runaway.
  • Universal Framework Interoperability: Zero vendor lock-in. Drop-in companion for LangGraph, LangChain, AutoGen, CrewAI, and custom raw Python loops, with native HTTP sidecar bridges for non-Python agents (Claude Code, OpenClaw, Hermes).

• snapjudge
  Fast, non-autoregressive System-1 decision engine in pure C++17.
  Answers typed questions (choice, score, noun) with calibrated probabilities in a single forward pass, delivering sub-millisecond execution with zero GPU dependencies.

• kibo-v7-
  Career orchestration platform.
  TypeScript, React 18, Vite, TanStack Query, PostgreSQL with Supabase Realtime sub-100ms CDC, Tailwind CSS, Shadcn UI, Recharts for garden graphs and leaderboards, v5.0.0 MIT.

Both CONTINUUM and SNAGLINE are community-driven, active open-source projects welcoming developers, researchers, and maintainers worldwide. Whether you want to add new recovery scenarios, optimize statistical anomaly heuristics, integrate new agent frameworks, or improve documentation and tests, contributions are warmly welcomed. Check out the open issues, RFCs, and contribution guidelines in each repository to get involved.

Engineering specialized, sample-efficient models designed exclusively for high-stakes operational domains, runnable on local or private infrastructure with zero external data egress.

• QwerySmith Model Family · [Research Paper (PDF)] · [GitHub Codebase]
  Production Text-to-SQL research pipeline and self-healing autonomous database agent (Qwen3-4B base foundation + QLoRA via Unsloth and TRL).
  MIT · Python 3.10+ · 88.5% In-Dist Acc · 55.7% Enterprise Acc (+43 wins vs 18 losses on Gretel, p=0.0019) · Zero-Shot Reasoning

  Addresses single-source catastrophic overfitting in domain-adapted LLMs by engineering a balanced multi-source curriculum (b-mc2 + gretelai), enforcing leak-proof split carving, and formalizing the Few-Shot Paradox. Ships with an interactive, autonomous CLI agent featuring schema introspection and execution-guided self-healing error recovery.

  • QwerySmith 2.0 (Multi-Seed Development):
    • QwerySmith-2.0-dev-seed1: Multi-seed training checkpoint (Seed 1) for enterprise SQL reasoning and variance control.
    • QwerySmith-2.0-dev-seed2: Multi-seed training checkpoint (Seed 2) for cross-split generalization.
    • QwerySmith-2.0-dev-seed3: Multi-seed training checkpoint (Seed 3) for ensembling and variance minimization.
  • QwerySmith 1.1:
    • LoRA Adapter: 132 MB parameter-efficient adapter weights trained on balanced multi-source curriculum.
    • Merged Model: Full 16-bit standalone model for vLLM, TGI, and Transformers.
    • GGUF Quantized Model: Quantized 4-bit (q4_k_m) model for local offline inference via Ollama and llama.cpp.
  • QwerySmith 1.0:
    • LoRA Adapter: 132 MB parameter-efficient adapter weights (single-source training).
    • Merged Model: Standalone 16-bit model checkpoint.
    • GGUF Quantized Model: Quantized 4-bit (q4_k_m) model for local offline inference.

• maya-healthcare-llm
  Research-grade sovereign healthcare operations LLM.
  Reproducible, production-oriented fine-tuning and evaluation suite specializing in non-clinical hospital administration, bed occupancy forecasting, workforce rostering, and operating room logistics.

• wave-coAtNet
  Wavelet-enhanced hybrid CNN-Transformer with 2D Discrete Wavelet Transform (DWT) subband decomposition.
  Features MBConv depthwise separable convolution, cross-attention, relative position bias, and prototype selection. Evaluated on 13 model benchmarks with 5-fold cross-validation, achieving 0.942 Quadratic Weighted Kappa across rare ichthyosis subtypes.

• H-CoAtNet-Ichthyosis
  Hierarchical hybrid CNN-Transformer for 5 ichthyosis subtypes.
  Benchmarked on 1,580 clinical images using PyTorch and timm, integrating convolutional stems with transformer blocks and squeeze-and-excitation modules.

• HI-MobileNet · [IEEE Publication] · [Codebase]
  MobileNetV2 architecture for Harlequin Ichthyosis classification.
  IEEE DOI 11332605, achieving 99.96% accuracy on clinical diagnostic subsets.

• H-CoAtNet-Ichthyosis-Classification
  Training and evaluation harness for WaveCoAtNet architectures.
  Supports 7 pretrained and scratch baselines with automated stratified validation splits.

• HAR from Video · [IEEE PICC 2025 Publication] · [Codebase]
  Sample-efficient 7-class human activity recognition benchmark.
  IEEE PICC 2025 DOI 11291542. Evaluated on 1,113 videos using CNN-LSTM architectures reaching 96.23% classification accuracy in Python, TensorFlow, and PyTorch.

• SPPS-Mac-Os-Code-Base
  Succinct Path Partition Serialization: Bijective O(n) tree serialization.
  Benchmarked on Apple M1 arm64 across 8 memory blocks versus LOUDS, FlatBuffers, and Protobuf. ESA 2026 Track E submission.

• spps-linux-experiment-results
  Cross-platform algorithmic validation on AMD EPYC 7763 Ubuntu 24.04.
  12,006 of 12,006 test suites verified passing (100% deterministic conformance).

• spps-experiments
  ESA 2026 submission benchmarks on AMD Ryzen 5 7235HS platform.

• tensorflow (197,312 stars, 76,097 forks)
  Core contributions in upstream TensorFlow across tensor math, gradient stability, and XLA compilers. Fork: Cyrax321/tensorflow (24 merged and upstream fixes).

  • fix-weighted-moments-tensor-axes: #122402 - Fix TypeError when axes is Tensor or ndarray. Normalize via constant_value and tolist. Add 3 tests.
  • fix-linalg-det-singular-gradient: #122823 - Fix singular matrix crash. Use SVD det times A inverse H plus pinv instead of matrix_inverse.
  • fix-topk-grad-int64-dtype: branch - Fix hardcoded int32 offset. Use dynamic index_type for top_k.
  • fix-grappler-argmax-saturating-ops: #122826 - Fix wrong results for saturating ops in float32.
  • fix-igamma-domain-nan: #124927 - Fix domain check for NaN boundary handling.
  View all 24 fixes

  • fix-weighted-moments-tensor-axes: #122402
  • fix-linalg-det-singular-gradient: #122823
  • fix-topk-grad-int64-dtype: branch
  • fix-grappler-argmax-saturating-ops: #122826
  • fix-igamma-domain-nan: #124927
  • fix-grappler-reciprocal-involution: #123195
  • fix-floordiv-negative-infinity: #123862
  • fix-resource-sparse-adagrad-dtype-mismatch: #124233
  • fix-xla-transpose-negative-perm: #124586
  • fix-xla-tensorarray-unstack-scalar: #124929
  • fix-numpy-cross-xla-static-shape: #124588
  • fix-mlir-reciprocal-involution: #123282
  • fix-igamma-grad-nan-boundary: #123803
  • fix-cumulative-logsumexp-nan: #115554
  • fix-speech-commands-exception-types: #117858
  • fix-speech-commands-python-idioms: #117894
  • fix-tools-exception-types: #117895
  • fix-generic-exceptions-python: #117860
  • fix-cmake-overridable-fetchcontent-doc: #124407
  • fix-doc-typos: #124450
  • fix-misc-typos: #115560
  • fix-posixpath-sys-path: #115551
  • fix/weighted-moments-tensor-axes: #122402
  • revert-115560-fix-misc-typos: branch

I actively experiment with state of the art foundation models, coding agents, and multi-agent workflows. My development methodology pairs agentic execution with human oversight: system architectures, invariant tests, and complex edge cases are designed, steered, and reviewed by me, while autonomous agents accelerate implementation, exploration, and refactoring loops. This human-in-the-loop workflow allows me to prototype rapidly, test novel research hypotheses, and maintain rigorous engineering standards across codebases.

Portfolio: https://anandhupshaji.me · LinkedIn: https://linkedin.com/in/anandhupshaji · Email: cyrax8590@gmail.com