Skip to content
View manjunathshiva's full-sized avatar
53 followers · 37 following

Achievements

Achievement: Pull Sharkx2Achievement: StarstruckAchievement: YOLO

Achievements

Achievement: Pull Sharkx2Achievement: StarstruckAchievement: YOLO

Block or report manjunathshiva

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
manjunathshiva/README.md
Manjunath Janardhan — a scale rail running from a 312K-parameter model on an ESP32-S3 up to 7B-plus on Apple Silicon, with live GitHub readouts

Microsoft LangChain NVIDIA Google Research Neo4j Azure AI Docs IBM   Python C MLX ESP32


I build agentic systems for enterprises by day, and spend my evenings finding out how little silicon a working model actually needs.

every card on this page is hand-built SVG, regenerated nightly from public GitHub data. no third-party widgets, no visitor counters.

01 · silicon

The through-line: take something that supposedly needs a datacenter, and land it on hardware you can hold.

Instrument readout: stars, public repos, commits, merged upstream pull requests Language spectrum by bytes across public repositories

femtoclaw · C
The world's smallest AI agent. A full tool-calling loop in pure C on a $4 ESP32, inside ~120 KB of RAM — no allocator games, no RTOS scheduler to hide behind.

turboquant-mlx · Python
MLX implementation of Google's TurboQuant: extreme weight and KV-cache compression for LLMs on Apple Silicon. The KV cache is the part everyone forgets, and it's what actually caps your context.

esp32-tinyllm · C
28.9M parameters doing interactive storytelling on an $8 ESP32-S3. Gemma-style per-layer embeddings keep 25M of those weights in flash, so the model outgrows the RAM it runs in.

esp32-gpio-llm · Python
312K parameters that turn plain English into GPIO commands. No WiFi, no cloud, no API key — the radio is off and it still works.

fmx · Python
A friendly CLI for Apple's on-device Foundation Model — fm chat / respond / schema on macOS 26, today.

fastcontext · Python
A preserved mirror of Microsoft's removed repo-exploration subagent (arXiv 2606.14066), patched to serve locally through mlx-lm. Research code shouldn't vanish because a repo got pulled.

02 · upstream

I go after the bugs that survive code review because everything still looks fine: a conversion boundary that silently drops a field. A good share of what's below is that same bug wearing different clothes — and each fix ships with the regression test that keeps it dead.

merged — 25 upstream

project what shipped pr
MicrosoftDocs/azure-ai-docs clarify declarative workflow response events #838
microsoft/agent-framework keep agent compaction config when HandoffBuilder clones participants #8329
microsoft/agent-framework parse Responses function_call_output so hosted tool results survive #8078
microsoft/agent-framework make concurrent FileCheckpointStorage saves stop racing #7757
microsoft/agent-framework keep cached + reasoning token counts through Foundry hosting #8334
ag-ui-protocol/ag-ui give .NET clients typed access to AG-UI client state #2752
microsoft/agent-framework keep AG-UI workflow reasoning in thread snapshots #8058
microsoft/agent-framework register built-in orchestration types so checkpoint restore stops failing #8258
microsoft/agent-framework stop dropping URL query parameters in the default HTTP request handler #7765
microsoft/agent-framework preserve tool call/result ordering when AG-UI splits a message #8005
microsoft/agent-framework surface AG-UI workflow intermediate events as reasoning #8003
microsoft/agent-framework keep parallel function_result contents through AG-UI conversion #7980
microsoft/agent-framework forward function_invocation_kwargs from DevUI to the agent #7779
microsoft/agent-framework preserve agent additional_properties when HandoffBuilder clones #7755
PrismML-Eng/Bonsai auto-fall back to MLX when a GGUF build is absent, instead of dying #123 · #126
NVIDIA/GenerativeAIExamples fix the ModelFilter ImportError that broke the examples on new huggingface_hub #208
neo4j-graphacademy fix the wrong fulltext_index_name in the GraphRAG hybrid retriever #6
langchain-ai/langchain default GITLAB_URL to gitlab.com instead of raising #14638

plus six smaller docs, README and dependency fixes across NVIDIA/GenerativeAIExamples, AIAnytime and collabnix/dockerbangalore.

in review — 11 signals out

project what i shipped pr
microsoft/agent-framework terminate chat completions SSE stream with data: [DONE] #8532
ag-ui-protocol/ag-ui use the canonical A2UI v0.9 basic catalog ID in .NET #2771
microsoft/agent-framework consolidate in-memory agent session storage in .NET #8417
microsoft/agent-framework validate declarative message properties; tighten strict skill script schemas #8119 · #8120
google-research/tabfm cast float64 targets before the device move; ship the safetensors extra #74 · #75
ibm-self-serve-assets/SuperKnowa repair the PDF retriever notebooks #22 · #23

03 · bench

Weekly contribution trace over the last 52 weeks

04 · what i tell teams

  • Run it on the smallest thing that works. A constraint you can't argue with is the fastest way to find out what your architecture actually needs.
  • The bug is usually at a boundary. Serialize, clone, convert, forward — that's where fields go missing, and where tests rarely look.
  • A fix without a regression test is a rumour. It'll come back under a new issue number.
  • Design patterns didn't stop mattering because the component is now a model. Agentic systems fail in ways software engineering already has names for.

05 · talks & workshops

Speaker and workshop author — sessions on agentic AI, MCP, and graph-backed retrieval, with the material kept public rather than locked in a slide deck.

aidevcon 2026 · PyData 2025 · Neo4j OSS 2025 · Agentic AI + MCP · ADK + MCP · DevFest Noida · AWS Community Day Mumbai

06 · reach me

Bring me something that isn't supposed to fit.

Website LinkedIn X

Pinned Loading

  1. femtoclaw femtoclaw Public

    The world's smallest AI agent. ESP32 + pure C. $4 chip, ~120KB RAM.

    C 21 3

  2. turboquant-mlx turboquant-mlx Public

    Extreme weight + KV cache compression for LLMs on Apple Silicon (MLX implementation of Google's TurboQuant)

    Python 84 23

  3. fmx fmx Public

    A friendly CLI for Apple's on-device Foundation Model — fm chat/respond/schema on macOS 26, today.

    Python 6 1

  4. esp32-tinyllm esp32-tinyllm Public

    Interactive on-device storytelling with a 28.9M-parameter LLM on an $8 ESP32-S3, using Gemma-style Per-Layer Embeddings to keep 25M parameters in flash.

    C 9 1

  5. esp32-gpio-llm esp32-gpio-llm Public

    A 312K-parameter language model that turns English into GPIO commands, running entirely on an ESP32-S3. No WiFi, no cloud, no API key.

    Python 6

  6. portfolio portfolio Public

    TypeScript 1