Sairam Ugge sairam0424

Sairam Ugge

GenAI & Backend Engineer @Ascendion · Multi-Agent LLM Orchestration · RAG · Event-Driven Backends

Python · Go · FastAPI · gRPC | Open-Source AI Infrastructure | 📍 Hyderabad, India

One AI-native developer-tooling stack, end to end — code intelligence (trelix, Graph-Forge), execution (ag-bash), orchestration (MindForge, ContextOS, Agent-Forge), and production safety (Tombstone, RateCap) — plus the LLM infrastructure underneath it (kelvran) and the MCP servers that connect it to agents (mcpsmiths).

🚀 About Me

🧠 GenAI & Backend Engineer @Ascendion (Jun 2024 – Present) — I build multi-agent LLM systems and event-driven backends that ship to real production traffic.
🤖 Co-built and production-hardened Pensieve, an AI process-orchestration engine — multi-agent LLM workflows, human-in-the-loop approval gates, real-time streaming, and governed LLM routing across cloud providers. (2K+ daily users across domains.)
🧩 AAVA Code (AI coding plugin for VS Code) — architected the backend and re-architected a single-agent prototype into a multi-agent orchestration system (Main-Agent + sub-agent via crewAI flows; 150+ skills, 40+ tools, ~60 commands); also contributed to the VS Code frontend. (3K+ daily users across 5+ client environments.)
🔁 Co-built a prompt-driven execution engine (LLM agent orchestration with RAG + ReAct) — raised first-pass acceptance from 65% to 85% and cut planning from 1.5h to 15min. (1.5K+ users.)
⚙️ Owned backend systems end-to-end (API design, data modeling, event-driven pipelines) — a decoupled Redis Streams + SSE pub/sub replaced client polling, cutting latency to sub-150ms. (10K+ events/day for 2.5K+ users.)
🎨 Co-built a GenAI wireframe generator (PRDs/sketches/prompts → production-ready UI artifacts) and a prompt-to-React system (wireframes → modular components + routing); also led a React → Angular re-architecture with an SSE-driven real-time backend that cut UI load latency ~30%.
🛠️ Open-source builder — author of MindForge, trelix, Tombstone, Graph-Forge (private), Agent-Forge (private), ContextOS, ag-bash & more — agent frameworks, code-intelligence engines, production-reliability tooling, and AI-native infrastructure spanning personal projects and the kelvran/mcpsmiths orgs. (See Featured Projects below.)
🏆 Competitive programmer — Google Code Jam '23 (AIR 420; 3,687 / 85,000+), Meta Hacker Cup '22, Flipkart GRiD '22 top tier; mentored 250–300 students in DSA.
🌱 Currently going deeper on multi-agent orchestration, RAG, and distributed-systems design.
📫 Reach me at uggesairam0000@gmail.com · 📄 Resume · 💼 LinkedIn

🔭 Now

Pushing trelix, ag-bash, and MindForge through their next hardening + release cycle
Growing kelvran's gateway/evals and mcpsmiths' MCP servers past solo-maintainer scale

🧩 Featured Projects

Open-source AI infrastructure I build in the open — agent frameworks, code-intelligence engines, and developer tooling. Commit counts are a build-signal only; each card describes architecture and tech, not adoption.

trelix, live — real terminal output, not staged: trelix query doing semantic search over its own retrieval-pipeline source, correctly ranking reciprocal_rank_fusion first for a query about RRF fusion.

🧠 Code Intelligence & Engines

Project	What it is & Tech
trelix · `1,453 commits`	Fast, reliable code intelligence — Tree-sitter AST parsing, contextual hybrid search, adaptive query planning, call-graph expansion & LLM synthesis across 20+ languages with zero infra. Ships as a PyPI CLI, a multi-arch Docker image (`ghcr.io/sairam0424/trelix`), MCP/LangChain/LlamaIndex retriever packages, plus a GitHub Action (`trelix-index-action`) for CI-based repo indexing. `Python` · `Tree-sitter` · `BM25` · `CLI`
gRPC Microservices · `316 commits` — Order Processing System	Polyglot event-driven microservices — gRPC internal RPC + REST gateway, etcd leader election, ACID inventory reservations, a Saga orchestrator, metric-based read routing across PostgreSQL replicas, Debezium/WAL CDC outbox, Bloom filters, Redis caching, DLQ + idempotency, and observability via Envoy L7 (OpenTelemetry/Prometheus/Grafana). `Go` · `Python` · `gRPC` · `PostgreSQL` · `Kafka` · `etcd`
Graph-Forge · `574 commits` (private)	AI-native distributed code-intelligence platform — code modeled as a Neo4j knowledge graph + Chroma semantic embeddings, served by ~19 polyglot microservices over gRPC/REST. Kafka ingestion, Tree-Sitter AST parsing, full OpenTelemetry/Jaeger/Prometheus/Grafana stack via Envoy. RAG over massive codebases. `Go` · `Python` · `gRPC` · `Neo4j` · `Chroma` · `Kafka` · `Next.js`
ag-bash · `510 commits`	AI-native bash interpreter with fork-speculation (`bash.fork()`/`bash.speculate()` — parallel speculative execution with copy-on-write branches, a capability no peer sandboxed-shell project has) — exposed as the `@ag-bash/bash` shell engine plus an MCP server and an agent terminal bridge. Tree-sitter WASM parser, esbuild (ESM+CJS), WASM runtimes (CPython, QuickJS, SQLite3). `TypeScript` · `WebAssembly` · `MCP` · `Tree-sitter` · `pnpm`

How trelix actually works (simplified from its own repo README — the real pipeline has more optional legs than this)

flowchart LR
    A[Repository] --> B[Parser: 26 languages]
    B --> C[Chunker + Embedder]
    C --> D[(Vector + SQLite index)]

    Q[User Query] --> E[Adaptive Router]
    E --> F[Vector Search]
    E --> G[Contextual BM25]
    E --> H[Graph Expansion]
    D --> F
    D --> G
    D --> H
    F --> I[RRF Fusion + Rerank]
    G --> I
    H --> I
    I --> J[LLM Synthesis / GraphRAG]

🛡️ Production Reliability & Delivery

Project	What it is & Tech
Tombstone · `974 commits`	Production intelligence layer for feature flags at scale — blast-radius gating, circuit-breaker auto-rollback, Merkle-linked audit log, causal incident correlation ("What Changed?"), and Knight Capital–style tombstoning of stale flags. Polyglot monorepo. `Go` · `Python` · `TypeScript` · `PostgreSQL` · `Redis`
RateCap · `347 commits`	Faithful open-source recreation of Stripe's four-tier rate-limiter and load-shedder architecture, built as a hybrid core-engine + sidecar system. `Go` · `API Gateway` · `Distributed Systems`

🤖 Agent Frameworks & Infrastructure

Project	What it is & Tech
MindForge · `1,786 commits`	Agentic-intelligence framework for Claude Code — v12.0.0 ships 221 slash commands, 164 specialized subagents, 354 skills, 216 personas, and 35 dynamic workflows, plus hooks, governance, cost-aware model routing, and true wave (parallel) execution. Streaming WebSocket SDK. Ships as `npx mindforge-cc` (7K+ monthly npm downloads). `Node.js` · `TypeScript` · `MCP` · `sql.js`
Agent-Forge · `209 commits` (private)	Framework-agnostic, self-improving AI agent infrastructure — a Karpathy-style propose/eval/score/commit-or-revert loop over a mutable markdown agent spec (`AGENT.md`), git-versioned, with an LLM-judge eval harness and held-out validation to guard against overfitting. `Python` · `FastAPI` · `pgvector` · `Celery` · `MCP` · `Docker`
ContextOS · `233 commits`	Intelligence layer for autonomous AI agents — SQLite/vector indexing, multi-agent orchestration, and resilience. Ships as core/CLI/MCP npm packages (`@context-os/*`) plus a spatial dashboard. `TypeScript` · `SQLite` · `vector search` · `MCP` · `React` · `Three.js`

🏢 Organization Projects — kelvran & mcpsmiths

Project	What it is & Tech
kelvran/gateway · `700 commits`	Unified AI infrastructure platform — LLM gateway with an embedded multi-layer cache, plus an independently-versioned `evals` subsystem (currently v0.10.1). `Go`
mcpsmiths/tracehub-mcp · `185 commits`	MCP server for querying OpenTelemetry traces across multiple observability backends (Jaeger, Tempo, Traceloop, Datadog) for LLM application debugging. `Python` · `MCP` · `OpenTelemetry`
mcpsmiths/cost-guard-mcp · `64 commits`	Pre-flight query cost and result-size guardrails for AI agents across BigQuery, Snowflake, and Databricks. `Python` · `MCP`

🛠️ Tooling & Lab

Project	What it is & Tech
CommandVault · `161 commits`	Universal AI command manager — browse/search/organize slash commands, skills, agents, plugins, rules, and hooks across Claude Code, Cursor, Copilot, Windsurf, and Aider. Indexes 350+ items via a VS Code extension, an 18-command CLI, and a three-tier search engine. `TypeScript` · `VS Code Extension` · `CLI` · `SQLite`
Inkforge · `103 commits`	AI-powered article generation system — notes/topic/code → human-readable Markdown articles, published to Dev.to, Hashnode, Medium & more. `TypeScript` · `Anthropic` · `AWS Bedrock`
Not-Humans-Lab	Umbrella docs tying together the Not-Humans-Lab suite: daily-dose (`156 commits`, AI-curated tech digest) and nh-deck (`58 commits`, Markdown slide-deck CLI), plus nh-skills (private). `TypeScript` · `Astro` · `CLI`
Not-Humans · `400 commits` (private)	Thin root workspace federating independent AI-infra sub-projects (SkillStack, Deep-Research, Agent-Hub, rate-limit-observatory) — polyglot Turborepo/pnpm + Next.js, Python (uv + Temporal + FastAPI), and Go services with a Qdrant vector store. `TypeScript` · `Python` · `Go` · `Temporal` · `Qdrant`

🌐 Portfolio & Edge Projects

Project	What it is & Tech
anvilry · `722 commits`	Engineering portfolio with four switchable experiences over one content source — SSG classic site, AI concierge chat (RAG-grounded, AWS Bedrock), WebGL Build Graph, and a keyboard-native developer terminal. Exposes a read-only MCP server. `Next.js 16` · `React 19` · `TypeScript` · `Tailwind v4` · `R3F`
Thunderboard-Labs · `38 commits`	On-device TinyML on the Silicon Labs Thunderboard Sense 2 (EFR32MG12, no NPU) — an 85%-accurate IMU gesture recognizer running fully on-chip at ~87.5ms. `Embedded ML` · `Cortex-M4` · `Edge Impulse`

🔌 MCP Servers

Five Model Context Protocol servers, spanning agent orchestration, shell execution, code search, distributed tracing, and query-cost guardrails:

Server	What it connects	Install
`mindforge-mcp-server`	The MindForge engine over MCP — knowledge graph, project health, audit log	`npx mindforge-mcp-server`
`@ag-bash/mcp-server`	AI-native sandboxed bash — agents get a real, observable shell	`npx @ag-bash/mcp-server`
`trelix-mcp`	Semantic code search for Claude Code, Cursor, Windsurf	`pip install trelix-mcp`
`tracehub-mcp`	Query distributed traces across Jaeger, Tempo, Traceloop, Datadog	`uvx tracehub-mcp`
`cost-guard-mcp`	Pre-flight cost/result-size guardrails for BigQuery, Snowflake, Databricks	`uvx cost-guard-mcp`

Full details (versions, downloads, license) are in the Open Source Packages table below; also browsable on the official MCP Registry.

🔥 Recent Highlights

MindForge v12.0.0 (2026-09-24) — first release cut for real external users; an independent 8-agent security/production audit closed 1 CRITICAL + 4 HIGH findings before the cutover.
trelix v3.4.1 (2026-09-27) — closed a CRITICAL symlink-following vulnerability that could leak local secrets to a third-party vision API, alongside shipping new raster-image indexing with LLM vision captioning (v3.4.0).
ag-bash (2026-09-25) — shipped a 9-phase security-hardening backlog to main and reached a fully green cross-platform CI matrix after root-causing deep Windows/macOS path-handling bugs.
Tombstone v2.0.1 (2026-09-09) — shipped after live E2E testing against a real Postgres/Kubernetes stack surfaced 10 real production bugs, including severe connection-pooling races.
ContextOS (@context-os/core v1.13.2) — shipped a deliberate breaking change eliminating an unfixable critical CVE (optional embedding peer-dependency) with in-band degradation reporting for MCP agents.
kelvran (2026-09-21 to 09-22) — 5 releases in a week (gateway v0.14.0→v0.14.2, evals v0.10.0→v0.10.1) fixing 2 CRITICAL + a HIGH-severity SSE error-swallowing bug found via back-to-back audits, alongside a live AWS Bedrock production pilot; gateway has since reached v0.15.0.
mcpsmiths/cost-guard-mcp — shipped a third query-cost engine (Databricks, 2026-09-13) and, the next day, closed a CRITICAL secret-redaction bug in a full-repo audit, raising coverage to 98.4%.
mcpsmiths/tracehub-mcp v0.5.0 (2026-09-14) — a real end-to-end dry run against a live Jaeger/OTel stack caught a CRITICAL silent-data-loss bug before it reached users; now at v0.12.2.
CNCF OpenFeature — contributed Tombstone's OpenFeature-compatible provider listing to the CNCF-affiliated OpenFeature project's interested-parties registry (merged 2026-08-18).

📡 Recent Activity

🎉 Merged PR #448 in sairam0424/trelix
💪 Opened PR #448 in sairam0424/trelix
🎉 Merged PR #447 in sairam0424/trelix
🎉 Merged PR #446 in sairam0424/trelix
🎉 Merged PR #445 in sairam0424/trelix

📈 Track Record

Auto-updated weekly by update-track-record.yml.

Metric	Value
Versioned releases shipped	174 across 11 projects
PRs merged (career total)	1,912
Repositories contributed to	64 (all-time, all contribution types)
Contributions (last 12 months)	10.5K+

🛡️ Security Posture

Live OSSF Scorecard scores — an automated, independent audit (branch protection, pinned dependencies, SAST, vulnerability history, and more) that requires zero setup on my end; it scans any public repo. Shown for every repo across my personal account and the kelvran/mcpsmiths orgs that OSSF has scanned so far, unfiltered — including the ones that score lower, since that's the honest picture:

Repo	OSSF Scorecard
kelvran/gateway
ag-bash
anvilry
mcpsmiths/cost-guard-mcp
Inkforge
RateCap
mcpsmiths/tracehub-mcp
trelix
MindForge
ContextOS
CommandVault

📦 Open Source Packages

Package	License	Build
`mindforge-cc` (npm)
`mindforge-mcp-server` (npm)		(same pipeline as mindforge-cc)
`@ag-bash/bash` (npm)
`@ag-bash/mcp-server` (npm)		(same monorepo as @ag-bash/bash)
`@ag-bash/agent-bridge` (npm)		(same monorepo as @ag-bash/bash)
`@context-os/core` (npm)	(registry `license` field is null — pending fix)	(CI omitted — no green run in 10+ attempts across ~11 days; no commits since 2026-09-17)
`@context-os/cli` (npm)	(same registry gap as @context-os/core)	(same pipeline as @context-os/core)
`@context-os/mcp` (npm)	(same registry gap as @context-os/core)	(same pipeline as @context-os/core)
`trelix` (PyPI)
`trelix-mcp` (PyPI)		(same release train as trelix)
`trelix-langchain` (PyPI)		(same release train as trelix)
`trelix-llama-index` (PyPI)		(same release train as trelix)
`tracehub-mcp` (PyPI)		(mcpsmiths org project)
`cost-guard-mcp` (PyPI)		(mcpsmiths org project)
`@commandvault/cli` (npm)		(also on the VS Code Marketplace)

Also distributed via: Homebrew (brew install sairam0424/tap/ag-bash, brew install sairam0424/tap/mindforge) · MCP Registry (mindforge-mcp-server, @ag-bash/mcp-server, trelix-mcp, tracehub-mcp, cost-guard-mcp).

Connect with me:

🛠 Tech Stack

Python	React	GitHub	Go	FastAPI	Flask	Django
Node.js	GraphQL	PostgreSQL	DynamoDB	MongoDB	Elasticsearch	MySQL
Redis	Kafka	Docker	Kubernetes	Terraform	AWS	Azure
Prometheus	Grafana	Linux	Nginx	Next.js	Angular	Tailwind CSS
JavaScript	Webpack	TypeScript	_Redux	_Playwright	_RabbitMQ	_Swagger
_WebSocket	_gRPC	_Git	_Postman	_Jira	_Jupyter	_Sass
_Firebase	_Supabase	_{Material UI}	_{ShadCN UI}	_Vue.js	_npm	_Hadoop
_Bun.js	_Vite	_NumPy	_Pandas	_TensorFlow	_Neo4j	_Cassandra
_Bash	_Loki	_GCP	_LangChain	_{Apache Spark}	_MATLAB	_Selenium

🏆 Competitive Programming

🥇 Google Code Jam 2023 — AIR 420 (3,687 / 85,000+)
🏅 Meta Hacker Cup 2022 — Global Rank 4,048 / 70,000+
🏅 Flipkart Grid 2022 — Rank 1,325 / 40,000+
👨‍🏫 Mentored 250+ students in Data Structures & Algorithms
🏅 GitHub Achievements: Pull Shark (gold), Pair Extraordinaire (gold), Quickdraw, YOLO

📊 GitHub Stats

✍️ Latest Writing

🏅 LeetCode Stats

📫 Connect With Me

💼 LinkedIn: https://linkedin.com/in/sairam0424
💻 GitHub: https://github.com/sairam0424
📧 Email: uggesairam0000@gmail.com

⚡ Philosophy

I build systems that scale, stream in real time, and survive production chaos. Multi-agent AI orchestration, event-driven backends, and open-source infrastructure are where I live — and competitive programming keeps the fundamentals sharp.

☕ Support Me

MindForge, trelix, ag-bash, and the other projects above are built and maintained solo, outside full-time work — if any of them save you time, a coffee is genuinely appreciated (never expected).

Navigation Menu