Good agents
need context.

I’m Himanshu. I build agentic AI, memory layers, and the backend systems they run on.

Explore my work I also explain what I build.

Agent sessionsource / evidenceEarlier memorystoredRevisionwaitingNext sessionready to recall

Original instructionflyctl deploy

One stored fact, with provenance.

Interactive illustration · no model calls

Memory & contextMCP & agent systemsBackend engineering

Built around the
hard parts.

Local and voice agents, data dependencies, scraper failures, and the context a model relies on. Six projects I want you to spend time with.

OmniDevLocal knowledge → agent action

On your MacFiles · Docs · ScreenshotsVision OCR

index locally, retrieve with citations

Knowledge engineOllama + SQLite Local

SwiftUI appAgent toolsMCP clients

Architecture illustration · Ollama mode

Local AI · Native macOSSwiftUI · FastAPI · Ollama · MCP

OmniDev

An agent that works with what’s on your Mac.

A native macOS app that searches files and screenshots with citations, edits code, and runs tests. Its local knowledge engine is also available to other agents through MCP.

Explore OmniDev Open the source

Inside the build

SwiftUI connects to a local FastAPI engine. Vision OCR makes screenshot text searchable, while Ollama embeddings and SQLite support file-grounded answers. The agent uses scoped tools, approval prompts, and restorable file snapshots. Ollama mode supports local inference; Gemini is an optional cloud provider.

BlastRadiusColumn-level lineage

Drop
user_id

Orders table

Buyer ID rename

Revenue dashboard

ML features

Trace the
downstream impact

Illustrative lineage · evidence comes from DataHub

Data infrastructurePython · DataHub · CLI · MCP

BlastRadius

See what a schema change breaks before it merges.

Traces a changed column through DataHub lineage to downstream tables, dashboards, and ML assets, then records the assessment alongside the data.

Explore the source Watch the demo

Inside the build

Separate column and table traversals distinguish evidenced impact from an unproven dependency. A bridge to dashboard input fields follows the final hop that lineage traversal misses. CLI checks can block a merge, while separate MCP tools scan changes and write findings back to DataHub.

FinEd Saathi

A voice-first financial literacy tutor for English and Hindi conversations, with a consented handoff to a tax specialist and memory for learning preferences.

LiveKit · Deepgram · Gemini · Murf · Voice agents

Explore the voice agent

ScrapeSentry

Detect silent scraper drift, quarantine broken output, and hold repairs for approval before canary and production checks. The public console replays recorded evidence.

Bright Data · FastAPI · MCP · Supabase

Try the recorded demo

Graft

Compile semantic HTML into typed WebMCP tools. Inspect the evidence, review the contracts, export an adapter, and verify what the deployed site exposes.

TypeScript · WebMCP · DOM compilation

Open the workbench

ContextFirewall

Audit an agent’s recalled context for stale facts, contradictions, secrets, and unsupported claims before assembling its context pack.

Cognee · FastAPI · MCP · Memory governance

Inspect the memory layer

More builds on GitHub

I care about what
happens after
the prompt.

The interesting work starts when an agent has to carry context across sessions, use a tool safely, or recover from a failed action.

I work across Python and TypeScript, APIs and data stores, MCP tools and user-facing products. I like making the state visible enough that someone else can inspect what happened.

Memory that can changeProvenance, revisions, forgetting, and retrieval.

Actions with boundariesScoped access, approval gates, and auditable decisions.

Backends that keep their promisesPersistent state, retry-safe writes, and observable workflows.

Written from
the work.

Technical articles, architecture decisions, and things I learned while building.

Observability · Build lessons

We instrumented an AI agent swarm with SigNoz, and its own telemetry told us we were wrong about almost everything

Memory · Architecture

I built a trust firewall for my AI agent’s memory — on Cognee’s four verbs

MCP · Security

Your MCP Server Says It Is Read-Only. Who Checked?

All articles on DEV

Let’s build
what’s next.

Agentic AI, memory infrastructure, or a backend problem you’re working through.

jhahimanshu653@gmail.com