I'm an ML Researcher focused on computer vision, temporal modeling and agent infrastructure. I build research that ships.

Areas: LLM Fine-Tuning & Text-to-SQL · Agent Infrastructure · Sample-Efficient HAR · Rare Disease Imaging · Hybrid CNN-Transformer

• HAR from Video - IEEE PICC 2025 - DOI 11291542 - ieee - 7 class HAR, 1,113 videos, CNN-LSTM 96.23% - Code: HAR-Sample-Efficient-Activity-Recognition
• HI-MobileNet - IEEE - DOI 11332605 - ieee - Harlequin Ichthyosis, MobileNetV2 99.96% - Code: H-CoAtNet-Ichthyosis-Models
• HAR-Sample-Efficient-Activity-Recognition - 7 class video HAR benchmark, Python, TensorFlow, PyTorch, 96.23% CNN-LSTM
• H-CoAtNet-Ichthyosis - Hierarchical hybrid CNN-Transformer for 5 ichthyosis subtypes on 1,580 images, Python, PyTorch, timm, conv stem plus transformer plus SE
• wave-coAtNet - Wavelet enhanced successor with cross attention and prototype selection, 13 model kappa, 5 fold CV
• H-CoAtNet-Ichthyosis-Classification - Training harness for WaveCoAtNet, 7 pretrained plus scratch baselines
• SPPS-Mac-Os-Code-Base - SPPS O(n) tree serialization on Apple M1 arm64, 8 blocks vs LOUDS, FlatBuffers, Protobuf, ESA 2026 Track E
• spps-linux-experiment-results - Cross platform validation on EPYC 7763 Ubuntu 24.04, 12006 of 12006 PASS
• spps-experiments - ESA 2026 submission on Ryzen 5 7235HS, bijective O(n)
• QwerySmith: Multi-Source Curriculum Fine-Tuning and Robust Evaluation for Text-to-SQL - 2026 - [Research Paper (PDF)] · [Code] - Qwen3-4B + QLoRA, 88.5% in-dist, 55.7% enterprise, self-healing SQL agent

I fine-tune, specialize, and evaluate open-source language models with a focus on sample efficiency, domain generalization, and deterministic SQL execution.

• QwerySmith Model Family · [Research Paper (PDF)] · [GitHub Codebase]
  Production Text-to-SQL research pipeline & self-healing autonomous database agent (Qwen3-4B base foundation + QLoRA via Unsloth & 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 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 & 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.

• 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 · ~2,240+ 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 & Environment Revalidation: Resumes from compact, versioned state contracts rather than lossy conversation dumps, enforcing staleness propagation across external dependency graphs before execution resumes.
  • Deny-by-Default MCP Server (12 Tools): Native drop-in integration for Claude Code, Cursor, and Anthropic toolchains with sub-second deterministic rollbacks.
  • Rigorous Battle Testing: 14 Phase 6 recovery scenarios 100% verified inside real Claude Code agent sessions; automated Docker builds on GHCR and interactive web demo.

• SNAGLINE · [Documentation] · [PyPI]
  Zero-dependency, microsecond-scale real-time failure detection & guardrails for LLM agents.
  MIT · Python 3.10+ · 13★ · 9 forks · 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).

• kibo-v7- - Career orchestration platform, TypeScript, React 18, Vite, TanStack Query, PostgreSQL with Supabase Realtime sub 100ms CDC, Tailwind plus Shadcn plus Recharts for Garden graph and leaderboard, v5.0.0 MIT

• tensorflow - An Open Source ML Framework - 197,312 stars, 76,097 forks. My fork Cyrax321/tensorflow has 25 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 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 a less equal 0 to not a greater than 0 for NaN handling

Portfolio: https://sx3svi1pkrbco9gt.vercel.app/ · LinkedIn: https://linkedin.com/in/anandhupshaji · Email: cyrax8590@gmail.com