const himanshu: Engineer = {
role: "Full-Stack AI Developer",
awards: "🏆 Won the Hugging Face Tiny Titan track with Thousand Tokens Theatre",
focus: ["AI Agents", "Full-Stack Systems", "Data Products", "Developer Tools"],
building: "Practical AI products with reliable APIs & polished interfaces",
backend: "FastAPI · Express · tRPC · Drizzle · PostgreSQL",
frontend: "React · Next.js · TypeScript · Tailwind",
edge: "Agent tool-calling with structured outputs",
mindset: "Production-minded. Idea → MVP → Ship. Fast.",
};| Area | Engineering focus |
|---|---|
| Agent systems | Tool calling, structured output, reliable degradation |
| Product delivery | FastAPI/Express backends with React/Next.js frontends |
| Data products | Multi-source analysis, dashboards, reproducible pipelines |
| Realtime UX | Streaming updates, Socket.IO/SSE flows, live status feedback |
| Open source | Fast iteration across forks, hackathons, and upstream ecosystems |
|
Pre-merge impact analysis for schema changes. Walks DataHub column-level lineage to find every downstream table, dashboard and ML asset a dropped column breaks, names each owner, then writes the verdict back into DataHub.
|
Policy-gated incident response for onchain treasuries. Revokes a risky ERC-20 approval only after a deterministic mandate check, transaction preflight and human approval. Local Gemma 4 reasoning stays advisory.
|
🖼️ framekindPrivate alt-text drafts that never leave the browser. YOLOS Tiny runs on-device with grounded detection overlays, shipped with a rerunnable Arm64 sweep showing the backend, not quantization, was the real lever.
|
OpenTelemetry instrumentation for multi-agent LLM systems, with importable SigNoz dashboards and alerts so agent runs are traceable like any other service.
|
💻 omnidevNative macOS AI dev cockpit. A local Gemma 4 model reads your files and screenshots, answers with citations, edits code and runs your tests. Fully offline, no API keys.
|
🛡️ railguardA policy and human-approval layer between an AI agent and the Razorpay API. Every write is executed, parked for review or denied by a named rule, with a hash-chained audit log.
|
🐝 devswarmMulti-agent app generator running on open models, wired so it watches itself. Every agent is traced into SigNoz, which makes a swarm run as debuggable as any ordinary service.
|
A trust firewall for AI coding-agent memory, built on Cognee. Audits every remembered fact for staleness, contradiction, secrets and evidence before it reaches your agent. MCP-native.
|
▸ blastradius: schema-change impact analysis
Core modules
- DataHub client walking column-level lineage hop by hop from the changed column
- Severity classifier separating breaking, risky and unaffected downstream assets
- Owner resolution for every impacted table, dashboard, job and ML asset
- Assessment written back into DataHub so the finding lives with the data
flowchart LR
PR["Schema change in a PR"] --> SCAN["blastradius scan"]
SCAN --> LIN["DataHub column-level lineage"]
LIN --> DOWN["Downstream tables, dashboards, ML assets"]
DOWN --> SEV["Severity and owner resolution"]
SEV --> BACK["Verdict written back to DataHub"]
▸ Circuit: policy-gated allowance revocation
Core modules
- Live allowance discovery from onchain Approval history, cross-checked across two RPCs
- Deterministic mandate check with canonical payload hashing
- Local Gemma 4 reasoning over loopback-only Ollama, advisory and unable to widen policy
- KeeperHub REST and MCP execution adapters behind one safety contract
flowchart LR
APPR["ERC-20 Approval event"] --> DISC["Allowance discovery across 2 RPCs"]
DISC --> POLICY["Deterministic mandate check"]
POLICY --> LLM["Gemma 4, advisory only"]
LLM --> HUMAN["Human approval, expiring single-use"]
HUMAN --> KH["KeeperHub execution, REST or MCP"]
KH --> PROOF["Independently verified zero allowance"]
▸ FrameKind: on-device alt text with a quality guardrail
Core modules
- YOLOS Tiny running in-browser through Transformers.js, image never leaves memory
- Grounded detection overlay turned into an editable spatial alt-text draft
- Rerunnable device sweep comparing FP32 and UINT8 across WASM and WebGPU
- Agreement guardrail scoring every configuration against the full-precision reference
flowchart LR
IMG["Image stays in browser memory"] --> MODEL["YOLOS Tiny via Transformers.js"]
MODEL --> DET["Grounded detections"]
DET --> ALT["Editable spatial alt-text draft"]
MODEL --> SWEEP["FP32 vs UINT8 sweep"]
SWEEP --> GUARD["Agreement guardrail vs reference"]
GUARD --> PICK["Fastest config that still agrees"]
| Track | Repositories |
|---|---|
| Best portfolio projects | blastradius, circuit-allowance-sentinel, framekind, otel-swarm, omnidev, railguard, devswarm, ContextFirewall |
| Agent products | dhurandhar, swarmshift, agentbench, sentinel-nosana-agent, prism-mistral-hackathon |
| Developer and automation tools | omnidev, vulcan, pr-review-agent, apivault-api-docs-generator, python-automation-training-toolkit, webmcptutorial |
| Observability | otel-swarm, devswarm, signoz-agent-observability |
| Onchain and ZK | circuit-allowance-sentinel, veilpledge, commitpass-stellar-rsvp, matchday-pulse |
| Full-stack products | FormOS, reporank, launchproof-ai, opportunity-scout, hireiq-recruiting-assistant, documate-ai |
| Data and notebooks | ipl-evolution-data-analysis, deeplearning, langchain-rag-tutorial-2026, financeiq-notion-finance-tracker, autopm-notion-product-manager |
| Open-source contributions and forks | langchain, smolagents, cognee, OpenMetadata, coral, hive |
Auto-regenerates daily via .github/workflows/snake.yml. Run it once from the Actions tab to seed the first frame.
I build open AI products and developer tools — in the open, for free.
If something I've shipped saved you time, sparked an idea, or you just want to back the next build, sponsorship keeps the work shipping and the tools free for everyone.
| Tier | Amount | You get |
|---|---|---|
| ☕ Espresso | $5 / mo |
My genuine thanks + your name in the supporters list below |
| ⚡ Builder | $15 / mo |
Above + early access to new tools and priority issue responses |
| 🚀 Backer | $50 / mo |
Above + a monthly 30-min talk — AI/agents, architecture, or your project |
| 🏆 Patron | $100+ / mo |
Above + your logo/link featured here and across my project READMEs |
Prefer a one-off? GitHub Sponsors supports one-time gifts too — every bit is appreciated.
🌟 Current Sponsors
- Email — jhahimanshu653@gmail.com
- X — x.com/jhahimanshu653
- LinkedIn — linkedin.com/in/himanshu748
- GitHub — github.com/himanshu748
